arXiv CSopen access

EchoSonar-R: A Multi-View Reasoning-Enabled Model for Disease Classification and Report Generation in Echocardiography

EchoSonar-R: A Multi-View Reasoning-Enabled Model for Disease Classification and Report Generation in Echocardiography

arXiv:2606.28164v1 [cs.CV] 26 Jun 2026

Darya Taratynova1 , Ahmed Aly1 , Numan Saeed1 Yaqub1 ⋆

⋆

, and Mohammad

Division of Computing and Mathematical Sciences, Mohamed Bin Zayed University of Artificial Intelligence (MBZUAI), Abu Dhabi, UAE {darya.taratynova, ahmed.aly, numan.saeed, mohammad.yaqub}@mbzuai.ac.ae

Abstract. Echocardiography is the most widely used non-invasive cardiac imaging modality, providing essential information for cardiovascular diagnosis. Interpreting an echocardiogram requires synthesizing complementary evidence across multiple heart views to identify abnormalities and produce structured clinical reports. While recent efforts focus on improving classification performance, most models lack explicit diagnostic reasoning and spatially grounded anatomical evidence, limiting clinician trust. We present EchoSonar-R, a multi-view reasoning-enabled vision–language model that jointly performs multi-label disease classification and report generation from echocardiography studies. EchoSonar-R combines a spatiotemporal video encoder with a structure-aware cardiac detector that provides spatially grounded anatomical cues to improve interpretability and clinician trust during cross-view reasoning. EchoSonar-R is trained in two stages: supervised fine-tuning (SFT) on reasoning-annotated targets, followed by Group Relative Policy Optimization (GRPO) with task-specific rewards that jointly align classification and report generation within a unified reinforcement-learning framework. Across a private multi-view dataset and two public benchmarks, EchoSonar-R improves macro balanced accuracy by 17.1% on the private set and 6.1% on MIMICEchoQA over the strongest baseline, achieves a GREEN clinical faithfulness score of 0.800, and produces interpretable reasoning traces grounded in multi-view visual evidence. Keywords: Echocardiography · Multi-view Reasoning · Reinforcement Learning · Vision Language Models (VLMs)

1

Introduction

Cardiovascular diseases (CVDs) are the leading cause of death worldwide, accounting for 19.8 million deaths in 2022, with over three-quarters of these deaths occurring in low- and middle-income countries (LMICs) [37,54]. The global CVD burden is rising due to population growth, aging, and increasing risk factors, 1 ⋆

Joint supervision.

2

Taratynova et al.

with projections of 35.6 million cardiovascular deaths annually by 2050 [22]. In routine clinical care, echocardiography is the most widely used non-invasive cardiac imaging modality [2, 7]. During a comprehensive echocardiographic study, cardiologists acquire and interpret multiple views synthesizing complementary anatomical and functional information to evaluate heart function, documenting their findings in structured clinical reports [2, 51]. Globally, echocardiography demand continues to grow, yet the supply of qualified sonographers is not keeping pace [53]. This disparity is particularly acute in LMICs, where access to echocardiography remains limited [3,39,48]. The growing gap between rising diagnostic demand and the global shortage of trained sonographers underscores the need for automated solutions capable of not only performing echocardiographic measurements but also synthesizing multi-view findings into explicit diagnostic reasoning and structured clinical reports, thereby improving the efficiency, consistency, and accessibility of expert-level cardiac assessment worldwide [27]. Automated echocardiographic analysis has gained momentum to reduce clinician workload and improve diagnostic consistency [12, 61]. Early deep learning approaches targeted single-task objectives such as view classification [17], chamber segmentation [25], and ejection fraction estimation [4, 20, 26, 28, 35]. While effective for isolated tasks, these models do not synthesize complementary information from the multiple views, thereby limiting both their diagnostic scope and clinical applicability. More recently, echocardiography foundation models such as EchoCLIP [11], EchoPrime [51], and PanEcho [16] have improved echocardiography representations via large-scale pretraining. However, these models primarily map visual inputs to classification labels or retrieve reports without generating explicit, cross-view grounded reasoning or producing clinician-style reports. This lack of interpretable reasoning limits trust and hinders adoption in clinical workflows [38]. This motivates a framework that fuses multi-view echocardiographic evidence, enables diagnostic reasoning, and generates comprehensive clinical reports. In this work, we introduce EchoSonar-R1 , the first multi-view reasoning-enabled model for echocardiographic disease classification and report generation (Fig. 1). Our contributions are: – a multi-view architecture that combines a view-aware video encoder with a structure-aware cardiac detector, enabling cross-view reasoning and providing spatially grounded anatomical cues alongside global motion dynamics for improved clinical trust. – a two-stage training pipeline: supervised fine-tuning (SFT) with reasoningannotated targets followed by group relative policy optimization (GRPO) with task-specific reward functions, that unifies multi-label classification and report generation within a single reinforcement learning (RL) framework. – extensive experiments on three datasets, demonstrating strong performance on both abnormality classification and report generation. 1

https://github.com/BioMedIA-MBZUAI/EchoSonar-R

EchoSonar-R

2

3

Related Work

Automated Echocardiography Report Generation. Automated report generation has progressed from early CNN-RNN image captioning frameworks [50, 55] to transformer-based architectures [49]. In radiology, advances span memorydriven transformers [9, 10], contrastive anomaly-focused generation [58], and knowledge graph grounding [21]. In echocardiography specifically, rather than generating text directly from images, existing approaches produce reports from structured measurement inputs provided by sonographers. Syryca et al. [44] evaluated ChatGPT’s ability to generate comprehensive echocardiography reports from numerical measurements, while Chao et al. [8] developed EchoGPT for echo report summarization. However, these methods depend on sonographer-derived measurements and cannot produce integrated reports directly from echocardiographic video. Medical Vision-Language Models. Contrastive vision–language pretraining [34] has driven rapid progress in medical VLMs. MedGemma [40] provides open medical foundation models with a medically tuned SigLIP encoder pretrained across diverse imaging modalities, while Lingshu [56] introduces a generalist medical MLLM spanning 12+ modalities and strong medical VQA performance. In ultrasound, EchoVLM [42] proposes a Mixture-of-Experts model trained across seven anatomical regions for report generation, diagnosis, and VQA. In echocardiography, EchoCLIP [11] aligns echo videos with clinical text via contrastive pretraining for zero-shot assessment, and its successor EchoPrime [51] extends this with a multi-view, study-level architecture for more comprehensive interpretation. Reasoning-Enabled Medical Models. Recent work has begun to incorporate explicit reasoning into medical VLMs, motivated by chain-of-thought (CoT) progress in general-purpose models [52]. MedVLM R1 [32] uses RL to encourage human-interpretable reasoning in medical image analysis, while Chiron-o1 [43] generates high-quality medical CoT data via curriculum learning, achieving strong results on medical VQA and reasoning benchmarks. However, these approaches largely focus on single-image or single-modality VQA and do not address the multi-view reasoning required in echocardiography. EchoSonar-R addresses this gap by enabling multi-view, reasoning-based report generation and disease classification, grounded in anatomical evidence.

3

Method

Given an echocardiographic study S = {V1 , V2 , . . . , VN } comprising N videos from distinct standard views and a task prompt q, EchoSonar-R generates a structured response Y = [⟨think⟩ R ⟨/think⟩ A], where R is a reasoning trace and A is the answer to q. As illustrated in Fig. 1, the model consists of three components: (1) a multiview visual encoder, processing each video Vi through a frozen spatiotemporal backbone and a structure-aware detector to extract complementary global and

4

Taratynova et al. A. Multi-View Visual Encoding

B. Cross-Modal Projection & Interleaving Projected Visual Tokens

Echocardiography Study, Visual Tokens, Video Token Projection MLP

Video Encoder D frozen Video 1

Detector

LV

LA RV

D frozen

...

% ,trainable

Structure Tokens, LVOT

Multi-View Sequence

TV RA

Example: MV

PLAX

Visual Tokens,

View Classifier

...

D frozen

A4C

... Projected Structure Tokens

D frozen

D frozen

LV

LA RV

LVOT

Structure Token Projection MLP

TV RA

MV

Generate a comprehensive echocardiographic report

...

C. Two-Stage Training

LLM

<think> Reasoning </think>

% trainable

Answer

Stage 2: GRPO

Multi-View Sequence

...

Structure features:

% ,trainable

Stage 1: SFT

Multi-View Sequence

Structure features:

A4C:

<|im_end|><|im_start|>assistant

Structure Tokens, Detector

<|im_start|>user PLAX:

Video Encoder

Video N

<|im_start|>system You are a medical imaging assistant specialized in echocardiography analysis.<|im_end|>

Ground truth, : <think> Reasoning </think> Autoregressive CE

Generation-Exploration, G completions

LLM

<think> ... </think>

D base + % LoRA

<think> ... </think> ... <think> ... </think>

D Frozen

% Trainable

LoRA adapted

Answer

Reward:

</think> gate

Output / GT

IoU or Cosine sim

length penalty

Reward computation

Fig. 1: Overview of EchoSonar-R. Multi-View Visual Encoding: each video is processed by a frozen spatiotemporal encoder and a frozen structure-aware detector to extract complementary global and anatomical tokens for seven cardiac structures. Cross-Modal Projection and Interleaving: visual and structure tokens are projected into the language model embedding space via trainable MLPs and interleaved with view identifiers to form a multi-view input sequence. Two-Stage Training: SFT fine-tunes the model on reasoning-annotated targets; GRPO subsequently optimizes a composite reward.

anatomical representations, (2) cross-modal projection layers, mapping the visual tokens into the language model embedding space and interleave them with view identifiers and the task prompt, and (3) a reasoning-enabled language model, synthesizing information across all views to produce Y. 3.1

Multi-View Visual Encoding

Spatiotemporal Video Encoder. Each video Vi ∈ S is processed by a frozen mViT-based encoder [51] to obtain a sequence of T = 393 visual tokens Zi ∈ RT ×dv . The encoder remains frozen throughout training to preserve its pretrained motion and view-specific priors, leaving cross-modal alignment to projectors. Unlike prior work that operates on single frames or individual views, this formulation retains the full token sequence from each video, allowing the downstream language model to attend to cues across the cardiac cycle. Structure-Aware Detector. While the spatiotemporal encoder captures global cardiac dynamics, echocardiographic interpretation further requires a localized understanding of individual anatomical structures. To this end, we train an RTDETR-L detector [63] to localize seven key cardiac structures. For each video Vi ,

EchoSonar-R

5

we run the detector across frames and extract the learned object query embed(i) dings ej ∈ Rds , corresponding to each detected structure j. Query embeddings for the same structure are averaged across frames within a video to obtain a (i) single per-structure representation ēj ∈ Rds , capturing the characteristic appearance of each anatomical region. This results in up to K structure tokens per video, providing the language model with spatially grounded anatomical cues, and allowing its outputs to be accompanied by corresponding bounding-box evidence. Cross-Modal Projection and Multi-View Interleaving. The visual representations from the spatiotemporal encoder and the structure-aware detector reside in different feature spaces and must be aligned with the language model’s embedding space. We introduce two learned projection modules: a video projection fv : Rdv → RdLM that maps each video token, and a structure projection fs : Rds → RdLM that maps each detector query embedding, to dLM hidden dimension of the language model. Both fv and fs are two-layer MLPs with GELU activation. We construct the input sequence by interleaving projected visual tokens with textual view identifiers, which are obtained by a pretrained view classifier. Concretely, for each video Vi acquired from view vi , we prepend a text token sequence encoding the view name, followed by the T projected video tokens and the projected structure tokens. 3.2

Stage 1: SFT

During SFT, the model learns to interpret the projected multi-view visual tokens, align them with clinical concepts, and generate structured text. This initialization ensures that RL begins with an informative policy operating over a well-calibrated visual input space, thereby stabilizing reward-based optimization [31, 41]. Reasoning Data Construction. For each training study S, we construct ground-truth target sequences Y ∗ = [⟨think⟩ R∗ ⟨/think⟩ A∗ ], pairing a reasoning trace R∗ with a reference answer A∗ . Each A∗ is derived from the groundtruth expert report, and R∗ is generated by prompting Qwen3-8B [59] with guideline-based templates [27] to produce a step-by-step diagnostic justification that logically leads to A∗ . To encourage diverse reasoning and prevent overfitting to a single output format, we construct training pairs across five complementary question types: (i) full report generation, (ii) conclusion-only generation, (iii) description of a specific cardiac structure, (iv) binary classification of a specific abnormality, and (v) multi-label classification. This multi-task formulation encourages the model to develop shared reasoning capabilities that generalize across clinical queries. Full details on the prompt templates and example question-answer pairs are provided in the Supplementary Material. Training Objective. We optimize the autoregressive cross-entropy loss over the target tokens, conditioned on the multi-view visual input and the task prompt: \mathcal {L}_{\text {SFT}} = -\sum _{t=1}^{|\mathcal {Y}|} \log \, p_\theta \!\left (y_t \mid y_{<t},\, \mathbf {X},\, q\right ),

(1)

6

Taratynova et al.

where X is the interleaved multi-view input sequence and θ denotes all trainable parameters. During SFT, we unfreeze both the cross-modal projection layers (fv , fs ) and the full language model parameters, while the visual encoders remain frozen. 3.3

Stage 2: Group Relative Policy Optimization (GRPO)

We apply GRPO [41] to move beyond token-level imitation of generated reasoning, which does not directly optimize for task-level correctness [57]. GRPO instead rewards completions based on their final clinical correctness, allowing the model to refine its decision boundaries through outcome-based exploration. For each training prompt x, we sample G independent completions {y1 , . . . , yG }. Each completion receives a scalar reward ri , from which we derive group-normalized advantages Âi = σrir−r̄ +ϵ , where r̄ and σr are the mean and standard deviation of rewards within the group. The policy is updated by maximizing the advantageweighted log-likelihood over completion tokens: \mathcal {L}_{\text {GRPO}} = -\frac {1}{G} \sum _{i=1}^{G} \hat {A}_i \cdot \frac {1}{T_i} \sum _{t=1}^{T_i} \log \pi _\theta (y_t \mid x, y_{<t}).

(2)

Following recent work on reasoning models [24, 60], we omit the KL divergence penalty from the standard GRPO objective. The KL term regularizes the policy toward the reference model, potentially constraining exploration [60]. Reward design. The composite reward balances structural validity, correctness, and output completeness: r = r_{\text {fmt}} \cdot r_{\text {cor}} + r_{\text {len}}.

(3)

Format reward, rfmt ∈ {0, 1}. The reward equals 1 if the completion contains </think> tag, and 0 otherwise. This gates the entire correctness signal, enforcing well-formed chain-of-thought structure as a prerequisite for receiving any credit. Correctness reward, rcor ∈ [0, 1]. The correctness reward rcor ∈ [0, 1] is taskdependent and designed to provide a smooth, differentiable training signal. (i) Multi-label classification. We compute: r_{\text {cor}}^{\text {multi-label}} = \frac {|\mathcal {P} \cap \mathcal {G}|}{|\mathcal {P} \cup \mathcal {G}|},

(4)

where P and G denote the predicted and ground-truth sets of abnormalities. IoU provides a graded signal that rewards partial correctness and penalizes both missed findings (false negatives in G \P) and hallucinated findings (false positives in P \ G) symmetrically. (ii) Report generation. We compute per-section semantic similarity using the pretrained echocardiography-aware text encoder: r_{\text {cor}}^{\text {report}} = \frac {1}{|\mathcal {C}|} \sum _{c \in \mathcal {C}} \cos \!\big (\mathbf {e}_c^{\text {pred}},\, \mathbf {e}_c^{\text {gt}}\big ) \label {eq:report_reward},

(5)

EchoSonar-R

7

where C is the set of report sections, and epred , egt c c are the text encoder embeddings of the predicted and ground-truth text for section c. Missing predicted sections are assigned a similarity of zero. We adopt semantic similarity rather than lexical metrics such as BLEU because echocardiographic reports allow substantial linguistic variation for clinically equivalent findings (e.g., “mildly dilated left atrium” vs. “LA is mildly enlarged”), which n-gram overlap metrics penalize despite identical clinical meaning. A specialized clinical text encoder better captures these domain-specific semantic equivalences. Length penalty, rlen ≤ 0. The length penalty discourages short completions:  min , where L is the completion length (tokens) and Lmin = rlen = min 0, L−L Lmin 400. Completions shorter than Lmin receive a penalty proportional to their deficit, while longer ones receive no penalty.

4

Experiments

EchoSonar-R is evaluated on two core tasks, abnormality classification and report generation, across three datasets spanning private multi-view and public single-view echocardiography: Private Dataset: We train and evaluate on a private echocardiography dataset with 5,061 training and 1,215 test studies, comprising 46,382 and 11,310 videos, respectively. Each study includes a median of 7 multi-view videos, with a median of 31 frames per video. The dataset covers 12 abnormality categories spanning valvular disease, chamber enlargement, systolic dysfunction, valve calcification, and structural anomalies, with prevalence ranging from 4.4% to 65.0%, reflecting natural class imbalance. Each report is structured into C=13 section-level descriptions followed by a clinical conclusion. For the structure-aware detector, a subset of 2,689 studies was annotated with segmentation masks for seven cardiac structures, converted to bounding boxes and propagated across frames using RAFT optical flow [46]. Full dataset statistics are provided in the Supplementary Material. MIMICEchoQA [47]: We additionally evaluate on MIMICEchoQA, a publicly available echocardiogram-based visual question-answering benchmark derived from the MIMIC-IV-ECHO database [13, 14]. The dataset comprises 622 clinician-validated, closed-ended QA pairs, each linked to a unique single-view echocardiogram video. We convert the closed-ended QA pairs into binary abnormality classification labels, retaining the subset of 155 samples pertaining to five of our target conditions. EchoNet-Dynamic [30]: We further evaluate on EchoNet-Dynamic, a large publicly available echocardiography dataset comprising A4C view videos collected during routine clinical care at Stanford University Hospital. Each video is paired with expert-annotated ejection fraction (EF) and left ventricular tracings. We use the standard test split of 1,277 studies and convert continuous EF values into a binary LV systolic dysfunction label: studies with EF in the range [55%, 70%] are labeled as normal, following established clinical guidelines that define normal LV systolic function within this range [19].

8

Taratynova et al.

Evaluation Metrics. We evaluate our model on two tasks: (i) abnormality classification, and (ii) report generation. Following prior work in medical image classification [36, 51], we report per-disease F1 scores, balanced accuracy, and macro-averaged metrics across all abnormality categories. We adopt balanced accuracy and macro-averaging rather than standard accuracy to account for the substantial class imbalance in our dataset, ensuring that performance on rare conditions contributes equally to the overall evaluation. To evaluate the quality of the reasoning traces produced within the <think> blocks, we employ an LLM-as-a-judge framework [64]. Each reasoning trace is scored on a 1–5 Likert scale across five dimensions: Reasoning–Answer Agreement, measuring whether every abnormality discussed in the reasoning trace is reflected in the final answer and vice versa; Reasoning Efficiency, assessing whether the reasoning is concise and focused, without redundant or circular steps; Factual Correctness, measuring the accuracy of clinical and anatomical claims made during reasoning; Evidence Grounding, evaluating whether the model references specific visual observations from the echocardiographic input rather than relying on generic statements; and Terminology Accuracy, assessing the correct use of echocardiographic and medical terminology throughout the reasoning chain. The generated reports are evaluated using natural language generation (NLG) metrics: BLEU-1/2/3/4 [33], which measure n-gram precision between generated and reference reports; METEOR [6], which extends n-gram matching with synonym recognition and stemming; and ROUGE-L [23], which captures the longest common subsequence between candidate and reference texts. However, these lexical overlap metrics are known to correlate poorly with clinical correctness [29], as semantically equivalent medical descriptions may use different surface forms. To address this, we compute BERTScore [62], which evaluates semantic similarity via token-level cosine similarities between contextual embeddings of candidate and reference reports. Furthermore, we report the GREEN score [29], which identifies and classifies clinically significant errors in generated reports. It produces a score between 0 and 1 that has been shown to align more closely with expert preferences than conventional NLG metrics [29]. Implementation Details. The spatiotemporal encoder is a frozen EchoPrime (mViT) [51], dv = 768; we also use its pretrained text encoder to compute the per-section semantic similarity reward in Eq. (5). The detector is RT-DETRL [63] trained for 30 epochs with class-balanced sampling, ds = 256, K = 7. Inputs are 16 frames per video with stride of 2. The language model is Qwen38B [59]. SFT is trained for 3 epochs (LR 2×10−5 , AdamW optimizer, batch size 64). GRPO uses G=8 completions (τ =1, top-p 0.95), LoRA (r=64, α=128) on all LM projections with frozen base weights, LR 1×10−6 , cosine schedule, 5 epochs, DeepSpeed ZeRO-2. BERTScore is computed using PubMedBERT [15]. Reasoning quality and GREEN score are computed using Mistral-7B [18].

EchoSonar-R

9

Table 1: Abnormality classification performance across private and public datasets. We report per-disease F1 score (↑) and balanced accuracy (BAcc, ↑). EchoSonar-R† denotes GRPO training; EchoSonar-R∗ denotes SFT-only. Baselines: Qwen3-VL [5], MedGemma [40], EchoVLM [42], Chiron-o1 [43], Lingshu [56]. Best results in bold, second best underlined. Prev. = prevalence (%).

Abnormality

Prev.

EchoSonar-R†

EchoSonar-R∗

Qwen3-VL

MedGemma

EchoVLM

Chiron-o1

Lingshu

F1

BAcc

F1

BAcc

F1

BAcc

F1

BAcc

F1

BAcc

F1

BAcc

F1

BAcc

66.6 68.3 64.9 38.0 59.5 42.3 71.6 45.4 57.4 25.9 13.8 39.2 49.4

59.9 61.6 75.8 62.3 74.0 65.4 82.2 72.2 75.5 58.7 55.5 65.7 67.4

66.6 69.1 62.7 36.2 55.9 41.3 69.8 44.5 53.9 18.4 3.6 19.2 45.1

59.3 62.6 74.6 61.3 71.9 65.1 80.9 69.7 71.5 55.9 50.6 57.4 65.1

38.3 49.7 26.6 20.1 21.8 24.2 14.9 15.4 11.2 6.0 2.5 4.5 19.6

52.1 51.5 49.8 49.9 51.8 52.3 50.3 51.7 52.5 46.9 44.2 50.1 50.3

27.9 24.6 35.1 23.7 19.6 16.8 10.0 9.4 8.8 9.1 2.3 6.7 16.2

51.2 49.2 49.3 48.1 48.6 50.4 49.7 48.1 47.9 50.9 41.1 55.6 49.2

0.0 0.0 0.0 0.0 0.0 26.8 5.8 0.0 0.0 0.0 0.0 0.0 2.7

50.0 50.0 50.0 50.0 50.0 50.0 50.9 50.0 50.0 50.0 50.0 50.0 50.1

61.1 69.6 40.2 0.0 27.9 26.1 16.6 14.1 11.1 8.9 5.4 5.2 23.9

48.8 50.5 49.5 49.9 50.5 50.0 50.9 47.7 51.1 48.4 53.8 51.9 50.3

62.3 33.6 29.5 1.9 28.0 25.6 18.5 15.9 5.0 8.7 5.2 5.1 19.9

49.4 50.5 48.0 50.4 51.2 49.9 54.7 51.2 46.2 48.0 52.1 51.4 50.3

11.6 38.3 23.8 18.2 20.0 22.4

52.7 64.0 60.1 60.6 57.5 59.0

6.1 31.1 22.7 8.3 31.6 20.0

45.8 60.3 59.7 52.1 62.5 56.1

58.3 53.3 30.0 52.6 27.6 44.4

37.5 49.1 58.8 46.4 51.1 48.6

66.7 0.0 30.0 18.2 0.0 23.0

32.1 50.0 58.8 55.0 50.0 49.2

0.0 80.6 0.0 0.0 0.0 16.1

50.0 50.0 50.0 50.0 47.6 49.5

66.7 65.5 78.9 72.0 27.0 62.0

53.6 44.9 66.8 52.1 47.1 52.9

86.7 38.1 62.9 44.4 28.6 52.1

58.9 37.9 50.5 41.4 55.1 48.8

60.2

37.2

58.7

5.6

50.6

51.2

50.5

66.7

50.0

36.0

51.8

7.7

50.0

Private Test Set (n = 1,215 studies, multi-view) TV Regurgitation MV Regurgitation LA Enlargement Healthy LV Systolic Dysfunction AV Regurgitation AV Stenosis MV Calcification LV Enlargement RA Enlargement RV Enlargement Bicuspid AV Macro Average

54.6 54.4 27.3 17.5 16.3 15.4 9.5 8.6 5.7 4.8 2.5 2.5

MIMICEchoQA [47] (single-view) LA Enlargement, n = 18 AV Regurgitation, n = 40 LV Systolic Dysfunction, n = 28 RV Enlargement, n = 17 AV Stenosis, n = 52 Macro Average

77.8 67.5 60.7 58.8 21.1

EchoNet-Dynamic [30] (n = 1,277, single-view) LV Systolic Dysfunction

4.1

39.3

Results of Abnormality Classification

Private test set. Tab. 1 presents per-disease classification results on the private multi-view test set. EchoSonar-R† demonstrates strong abnormality classification performance, outperforming all baselines across both F1 and BAcc. For macro BAcc, EchoSonar-R† achieves 67.4%, surpassing the SFT-only EchoSonarR∗ by 2.3% and the strongest external baselines (Chiron-o1 and Lingshu) by over 17%. For macro F1, EchoSonar-R† achieves 49.4%, approximately twice the best external baseline (Chiron-o1, 23.9%). The advantage is particularly pronounced on low-prevalence conditions: for aortic stenosis (9.5%), EchoSonar-R† attains 82.2% BAcc compared to 54.7% for the strongest baseline; for LV enlargement (5.7%), 75.5% versus 52.5%; and for bicuspid aortic valve (2.5%), 65.7% versus 55.6%. Notably, most external baselines hover around 50% BAcc across conditions, suggesting a drift toward majority-class predictions, resulting in reasonable F1 on common diseases while demonstrating poor separation between positive and negative cases. Comparing EchoSonar-R† and EchoSonar-R∗ , EchoSonarR† provides consistent improvements over SFT-only, with macro F1 rising from 45.1% to 49.4% and BAcc from 65.1% to 67.4%. The gains span 11 of 12 categories and are most pronounced for low-prevalence conditions, suggesting that reward-based optimization is particularly effective where SFT has a limited supervision signal. To assess statistical significance, we perform a per-class sign test: EchoSonar-R† outperforms EchoSonar-R∗ on F1 in 11 of 12 categories and on BAcc in 11 of 12, yielding a sign-test p ≈ 0.006, confirming that the GRPO

10

Taratynova et al.

Table 2: Reasoning quality evaluation on the private test set, scored by Mistral-7B [18] on a 1–5 Likert scale (↑). EchoSonar-R† denotes GRPO training; EchoSonar-R∗ denotes SFT-only. Best results are in bold, second best underlined. EchoSonar-R†

EchoSonar-R∗

Qwen3-VL

MedGemma

Chiron-o1

Lingshu

Reasoning-Answer Agreement Reasoning Efficiency Factual Correctness Evidence Grounding Terminology Accuracy

4.98 4.98 4.99 5.00 5.00

4.96 4.97 5.00 5.00 5.00

3.29 3.73 4.77 3.31 4.96

3.33 3.74 4.15 3.71 4.29

3.50 3.80 3.49 3.56 3.69

4.02 4.58 4.29 4.68 4.70

Average

4.99

4.99

4.01

3.84

3.61

4.45

Metric

gains are statistically significant. This pattern suggests that GRPO refines the model’s decision boundaries rather than learning fundamentally new visual features. For the MV Regurgitation where EchoSonar-R∗ outperforms EchoSonarR† , the difference is modest (<1% BAcc), indicating that EchoSonar-R† does not degrade performance on well-learned categories while substantially improving others. Public test sets. On MIMICEchoQA in Tab. 1, EchoSonar-R† achieves the highest macro BAcc of 59.0%, outperforming the Chiron-o1 by 6.1%. While baselines such as Chiron-o1 and Lingshu achieve higher macro F1, their BAcc remains at or below 52.9%. EchoSonar-R† maintains the highest BAcc on 3 of 5 conditions, being robust to cases containing single-view inputs. For aortic stenosis, EchoSonar-R∗ achieves the highest BAcc of 62.5%, outperforming Lingshu by 7.4%. On EchoNet-Dynamic, EchoSonar-R† achieves the highest BAcc of 60.2% for LV systolic dysfunction, surpassing the Chiron-o1 by 8.4%. Cross-dataset consistency. Five conditions are shared between the private test set and MIMICEchoQA, allowing direct comparison of model robustness across dataset shift. EchoSonar-R† exhibits the 8.4% BAcc average drop across these conditions. AV regurgitation is the most stable for EchoSonar-R, while LA enlargement shows the largest drop. This likely reflects both the difficulty of assessing chamber size from a single view and differences in diagnostic thresholds across institutions: trace or borderline findings may be reported as abnormal at one site but considered within normal limits at another. For LV systolic dysfunction, which can be evaluated across all three datasets, EchoSonar-R† maintains strong performance: 74.0% on the private set, 60.1% on MIMICEchoQA, and 60.2% on EchoNet-Dynamic. Reasoning quality. Tab. 2 reports reasoning trace quality across five dimensions. EchoSonar-R† achieves near-perfect average scores (4.99/5.00), substantially outperforming all baselines. The gap is widest on Evidence Grounding, indicating that baseline reasoning traces rely on generic medical statements rather than referencing specific visual observations from the echocardiographic input. Reasoning–Answer Agreement shows a similar pattern: EchoSonar-R† maintains near-perfect consistency between its reasoning and final answers (4.98), while Qwen3-VL (3.29) and MedGemma (3.33) frequently produce conclusions that contradict or omit findings discussed in their reasoning. Among baselines, Ling-

EchoSonar-R

11

Table 3: Report generation quality on the private test set. NLG metrics measure lexical and semantic overlap with reference reports (↑). GREEN [29] evaluates clinical faithfulness by identifying and categorizing clinically significant errors in generated reports. EchoSonar-R† denotes GRPO training; EchoSonar-R∗ denotes SFT-only. Best results are in bold, second best underlined. Metric

EchoSonar-R†

EchoSonar-R∗

Qwen3-VL

MedGemma

EchoPrime

Chiron-o1

Lingshu

0.795 0.767 0.740 0.725 0.829 0.819 0.985

0.792 0.762 0.736 0.720 0.826 0.815 0.985

0.076 0.027 0.014 0.010 0.195 0.113 0.924

0.135 0.066 0.047 0.038 0.260 0.188 0.929

0.143 0.061 0.036 0.027 0.234 0.192 0.933

0.125 0.039 0.023 0.019 0.197 0.185 0.931

0.099 0.031 0.016 0.011 0.230 0.162 0.927

0.796 0.602 0.179 0.198 0.019

0.216 1.747 0.706 0.677 0.044

0.453 1.275 0.260 0.716 0.027

0.306 1.588 0.524 0.585 0.047

0.358 1.502 0.475 0.613 0.044

0.491 1.193 0.321 0.633 0.015

NLG Metrics (↑) BLEU-1 BLEU-2 BLEU-3 BLEU-4 METEOR ROUGE-L BERTScore

GREEN Clinical Faithfulness GREEN Score (↑) Mean Sig. Errors (↓) Hallucination (↓) Omission (↓) Wrong Location (↓)

0.800 0.590 0.175 0.193 0.018

shu achieves the highest average (4.45), driven by strong Terminology Accuracy (4.70) and Evidence Grounding (4.68), yet its classification performance remains near chance (Tab. 1). 4.2

Results of Report Generation

Tab. 3 presents report generation quality on the private test set. EchoSonar-R† achieves the highest scores across all NLG metrics by a wide margin. For BLEU4, it reaches 0.725 compared to 0.010 for the Qwen3-VL, reflecting the advantage of training on reports from the same institutional format. Baselines that have never seen reports from this institution cannot match its lexical structure, making direct NLG comparisons with external models inherently asymmetric. BERTScore partially mitigates this by measuring semantic rather than surfacelevel overlap, yet EchoSonar-R† still leads (0.985 vs. 0.933) for EchoPrime and 0.931 for Chiron-o1, suggesting that the advantage extends beyond formatting to clinical content. The GREEN evaluation provides a more meaningful comparison, as it assesses clinical faithfulness independently of reporting style. EchoSonar-R† achieves a GREEN score of 0.800, substantially above Lingshu (0.491) and Chiron-o1 (0.358). The error breakdown reveals that Qwen3-VL produces 1.747 mean significant errors per report, with omissions (0.677) and hallucinations (0.706). It frequently fabricates findings not present in the study or omits pathology. Lingshu, on the other hand, has the highest omission rate (0.633) but comparatively lower hallucination (0.321), suggesting it tends toward underreporting rather than fabrication.EchoSonar-R† maintains the lowest error rates across all categories, with hallucination at 0.175 and omission at 0.193, and achieves the fewest wrong location errors (0.018).

12

Taratynova et al.

Table 4: Ablation studies on the private test set. Top: effect of reasoning-augmented training data during SFT. Bottom: contribution of visual encoder components at inference. VT = video tokens, ST = structure tokens. Best results per section are in bold. Macro Metrics Configuration

VT ST

F1

BAcc Sens. Spec.

✓ ✓

45.1 44.4

65.1 65.5

43.4 47.6

86.8 83.4

65.1 63.9 49.8 50.2

43.4 41.9 11.5 10.3

86.8 85.8 88.0 90.2

Reasoning in Training w/ reasoning targets w/o reasoning targets

✓ ✓

Visual Input Components at Inference Full Remove structure tokens Remove video tokens Remove both

✓ ✓ ✗ ✗

✓ ✗ ✓ ✗

45.1 43.9 10.5 10.2

Comparing EchoSonar-R† and EchoSonar-R∗ variants, the differences are modest: GREEN score improves marginally from 0.796 to 0.800, with small reductions in hallucination and omission. This is expected, as the GRPO reward for report generation uses cosine similarity, which optimizes for semantic content coverage rather than fine-grained error reduction. The SFT stage, trained directly on reference reports, already captures most of the institutional reporting conventions, leaving limited room for GRPO to improve further on this task. 4.3

Ablation Analysis

Effect of reasoning in training. Tab. 4 compares EchoSonar-R∗ trained with and without reasoning targets. Aggregate performance is comparable (macro F1: 45.1% vs. 44.4%; BAcc: 65.1% vs. 65.5%), suggesting that reasoning supervision does not substantially affect overall discriminative accuracy at SFT stage. The main change is in the sensitivity–specificity trade-off: removing reasoning traces increases sensitivity (47.6% vs. 43.4%) at the expense of specificity (83.4% vs. 86.8%). Thus, the reasoning-trained model behaves more conservatively, producing fewer positive predictions and fewer false positives, while the non-reasoning model identifies more true positives, but with more false positives. However, the reasoning traces support clinical transparency: they enable clinicians to audit the model’s stated rationale and more easily localize systematic failure modes than is possible with binary outputs alone. Visual input components at inference. Tab. 4 evaluates the contribution of each token type by removing them from the input sequence X at inference. The video tokens are the primary information source: removing them causes a significant drop in F1 (45.1% to 10.5%) and BAcc (65.1% to 49.8%), with the model falling to near-random discrimination. Removing both token types confirms this, with performance collapsing further to 10.2% F1 and 50.2% BAcc. Notably, the marginal difference between removing video tokens alone and removing both indicates that structure tokens carry negligible diagnostic signal without the accompanying video context. The structure tokens provide a com-

EchoSonar-R

13

Linear Probe AUROC: Projected vs. Raw Features 100

+1.5%

AUROC (%)

90

+3.5% +0.9%

80

+2.9%

-0.4%

+2.9%

+2.2%

After Projection Before Projection

+0.8%

+2.9%

+2.5%

+1.9%

-0.4%

70 60 50

TV Regurg.

MV Regurg.

LA Enlarg.

Healthy

LV Syst. Dysf. Aortic Regurg. Aortic Stenosis

MV Calcif.

LV Enlarg.

RA Enlarg.

RV Enlarg.

Bicuspid AV

Fig. 2: Linear probe AUROC before and after cross-modal projection.

plementary but smaller contribution: removing them while retaining video tokens produces a modest decrease in both F1 (45.1% to 43.9%) and BAcc (65.1% to 63.9%), confirming that the detector embeddings help calibrate predictions and improve discrimination between positive and negative cases. The rising specificity as tokens are removed (86.8% to 88.0% to 90.2%) reflects a progressively more conservative model that defaults to predicting the negative class when deprived of visual evidence. Effect of cross-modal projection. To test whether the learned projectors fv enhance the clinical separability of visual features, we train linear probes on EchoPrime embeddings before and after projection and report per-disease AUROC in Fig. 2. Projection improves AUROC for 10 of 12 conditions, with the largest gains for LV systolic dysfunction (3.5%), RV enlargement (2.9%), TV regurgitation (2.9%), and MV calcification (2.9%). The remaining two conditions, RA enlargement and bicuspid AV, show only minor decreases. Overall, these results indicate that fv reshapes the frozen encoder’s representation space to amplify the disease-relevant signal, acting as a learned task-specific adapter. 4.4

Qualitative Results

Fig. 3 illustrates representative model outputs. Panel (a) shows a normal study in which EchoSonar-R† correctly identifies the absence of disease by systematically reviewing each cardiac structure across five views and explicitly cross-checking consistency across views (e.g., verifying that normal LA dimensions in A4C agree with the PLAX assessment). Lingshu attains an average reasoning score of 4.2, yet it hallucinates pathologies with convincing clinical language, describing “turbulent flow across the tricuspid valve” and “thickened and calcified” aortic leaflets in a structurally normal heart. Chiron-o1, on the other hand, outputs a flat list labeling all 11 conditions as present with no reasoning or visual evidence. The EchoSonar-R∗ variant produces one false positive, which EchoSonar-R† learns to suppress as diastolic dysfunction. Panel (b) presents a more challenging case with multiple mild abnormalities present at different sections of the report. EchoSonar-R† correctly identifies

14

Taratynova et al.

(a) Classification: Normal study, ground truth = no abnormalities EchoSonar-R† : 0/11 errors, score: 5.0 R. [PLAX] LV normal. AV normal morphology, no calcification. MV normal mobility. [A4C] LA within normal limits, consistent with PLAX. RA normal. RV-to-LV ratio normal. [A4C Zoom] Wall motion normal, confirming systolic function from PLAX. [A2C] Wall motion normal. LA normal across views. [PSAX] No PR jet. [A3C] AV, MV confirmed normal across all views. A. No significant abnormalities. ✓

Lingshu: 10/11 errors, score: 4.2 Turbulent flow across TV, indicating regurgitation. MR as seen by backflow into LA. AV appears thickened and calcified. MV leaflets show calcification. RA appears enlarged. LV shows reduced contractility.

TR, MR, AS, MV Calc., RAE, RVE, LAE, LVSD, LVE ✗

EchoSonar-R∗ (SFT): 1/11 errors, score: 5.0 Chiron-o1: 11/11 errors, score: 1.4 R. [A4C Zoom] LV normal. LA enlarged, likely di- Based on the analysis of the echocardiographic astolic dysfunction. RV normal. [A2C] LA en- images, the identified abnormalities are: TR: larged. [PLAX] LV, MV normal. AV trileaflet. Present. MR: Present. AS: Present. MV Calc.: Present. RAE: Present. RVE: Present. LAE: Present. LVSD: Present. LVE: Present. BAV: Present. AR: Present. A. Left Atrial Enlargement ✗ (FP) All 11 abnormalities “Present” ✗ (b) Report generation: Per-section comparison Section Reference

EchoSonar-R†

Chiron-o1

Lingshu

LV

Mild concentric LVH. Normal systolic function. Normal appearance. Trace MR. Mild LVH. Normal EF.

Normal size and function.

Normal, no hypertrophy.

Normal, no regurgitation. Normal, no valvular disease.

Normal, no regurgitation. Normal, no abnormalities.

MV Concl.

Mildly reduced systolic function. Mild concentric LVH. Normal appearance. Mild MR. Mild hypokinesis. Mild LVH. Mild LAE. Mild MR.

Fig. 3: Qualitative examples. correct/grounded, cross-view confirmation, severity error, fabricated/missed.

mild concentric LVH but underestimates the severity of systolic dysfunction. In contrast, Chiron-o1 and Lingshu both report entirely normal findings across all sections, with no evidence of LVH, LA enlargement, or mitral regurgitation. This pattern is consistent with their near-chance BAcc in Tab. 1 and high omission rates in Tab. 3.

5

Conclusion

In this paper, we present EchoSonar-R, a reasoning-enabled multi-view vision– language model for echocardiographic disease classification and report generation. By integrating a spatiotemporal video encoder with a structure-aware cardiac detector, our model provides the language model with both global motion dynamics and spatially grounded anatomical cues across the full echocardiographic study. A two-stage training strategy enables the model to produce reasoning while optimizing directly for clinical correctness. Our extensive experiments across a private multi-view dataset and two public benchmarks demon-

EchoSonar-R

15

strate strong performance on both abnormality classification and report generation, improving macro balanced accuracy by +17.1% and +6.1% over the strongest baselines on private and public sets, respectively, while halving the mean clinically significant errors in generated reports (0.590 vs. 1.193), with ablations and qualitative analysis validating each component. Limitations and Future Work. The current model processes each study independently without leveraging prior examinations from the same patient. We plan to expand EchoSonar-R to incorporate longitudinal patient history, enabling tracking of disease progression across serial echocardiograms.

Acknowledgements The authors would like to thank Roman Sendler and Joseph Sokol from iCardio for their valuable feedback and insight.

References 1. Abdin, M., Aneja, J., Behl, H., Bubeck, S., Eldan, R., Gunasekar, S., Harrison, M., Hewett, R.J., Javaheripi, M., Kauffmann, P., et al.: Phi-4 technical report. arXiv preprint arXiv:2412.08905 (2024) 22 2. Ahmed, I., Sasikumar, N.: Echocardiography imaging techniques. In: StatPearls [Internet]. StatPearls Publishing (2023) 2 3. Arega, T., Desai, M.Y., et al.: Comparison of cardiovascular imaging practices in Africa, North America, and Europe: Two faces of the same coin. European Heart Journal – Imaging Methods and Practice 1(1), qyad005 (2023). https://doi.org/ 10.1093/ehjimp/qyad005 2 4. Asch, F.M., Poilvert, N., Abraham, T., Jankowski, M., Cleve, J., Adams, M., Romano, N., Hong, H., Mor-Avi, V., Martin, R.P., et al.: Automated echocardiographic quantification of left ventricular ejection fraction without volume measurements using a machine learning algorithm mimicking a human expert. Circulation: Cardiovascular Imaging 12(9), e009303 (2019) 2 5. Bai, S., Chen, K., Liu, X., Wang, J., et al.: Qwen3-VL technical report. arXiv preprint arXiv:2511.21631 (2025) 9, 27 6. Banerjee, S., Lavie, A.: METEOR: An automatic metric for MT evaluation with improved correlation with human judgments. In: Proceedings of the ACL Workshop on Intrinsic and Extrinsic Evaluation Measures for Machine Translation and/or Summarization. pp. 65–72 (2005) 8 7. Chambers, J.: Echocardiography: Frontier imaging in cardiology. British Journal of Hospital Medicine (2012), pMC3473911 2 8. Chao, C.J., Banerjee, I., Arsanjani, R., Ayoub, C., Tseng, A., Delbrouck, J.B., Kane, G.C., Lopez-Jimenez, F., Attia, Z., Oh, J.K., Erickson, B., Fei-Fei, L., Adeli, E., Langlotz, C.: Evaluating large language models in echocardiography reporting: Opportunities and challenges. European Heart Journal – Digital Health 6(3), 326– 339 (2025). https://doi.org/10.1093/ehjdh/ztae086 3 9. Chen, Z., Shen, Y., Song, Y., Wan, X.: Cross-modal memory networks for radiology report generation. arXiv preprint arXiv:2204.13258 (2022) 3

16

Taratynova et al.

10. Chen, Z., Song, Y., Chang, T.H., Wan, X.: Generating radiology reports via memory-driven transformer. In: Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP). pp. 1439–1449 (2020) 3 11. Christensen, M., Vukadinovic, M., Yuan, N., Ouyang, D.: Vision–language foundation model for echocardiogram interpretation. Nature Medicine 30(5), 1481–1488 (2024) 2, 3 12. Ghorbani, A., Ouyang, D., Abid, A., He, B., Chen, J.H., Harrington, R.A., Liang, D.H., Ashley, E.A., Zou, J.Y.: Deep learning interpretation of echocardiograms. npj Digital Medicine 3, 10 (2020). https://doi.org/10.1038/s41746-019-0216-8 2 13. Goldberger, A.L., Amaral, L.A.N., Glass, L., Hausdorff, J.M., Ivanov, P.C., Mark, R.G., Mietus, J.E., Moody, G.B., Peng, C.K., Stanley, H.E.: PhysioBank, PhysioToolkit, and PhysioNet: Components of a new research resource for complex physiologic signals. Circulation 101(23), e215–e220 (2000) 7 14. Gow, B., Pollard, T., Greenbaum, N., Moody, B., Johnson, A., Herbst, E., Waks, J.W., Eslami, P., Chaudhari, A., Carbonati, T., et al.: Mimic-iv-echo: Echocardiogram matched subset (2023) 7 15. Gu, Y., Tinn, R., Cheng, H., Lucas, M., Usuyama, N., Liu, X., Naumann, T., Gao, J., Poon, H.: Domain-specific language model pretraining for biomedical natural language processing. ACM Transactions on Computing for Healthcare 3(1), 1–23 (2022). https://doi.org/10.1145/3458754 8 16. Holste, G., Oikonomou, E.K., Tokodi, M., Kovács, A., Wang, Z., Khera, R.: Panecho: Complete ai-enabled echocardiography interpretation with multi-task deep learning. medRxiv pp. 2024–11 (2025) 2 17. Jansen, G.E., de Vos, B.D., Molenaar, M.A., Schuuring, M.J., Bouma, B.J., Išgum, I.: Automated echocardiography view classification and quality assessment with recognition of unknown views. Journal of Medical Imaging 11(5), 054002–054002 (2024) 2 18. Jiang, A.Q., Sablayrolles, A., Mensch, A., Bamford, C., Chaplot, D.S., de las Casas, D., Bressand, F., Lengyel, G., Lample, G., Saulnier, L., Lavaud, L.R., Lachaux, M.A., Stock, P., Le Scao, T., Lavril, T., Wang, T., Lacroix, T., El Sayed, W.: Mistral 7B. arXiv preprint arXiv:2310.06825 (2023) 8, 10 19. Lang, R.M., Badano, L.P., Mor-Avi, V., Afilalo, J., Armstrong, A., Ernande, L., Flachskampf, F.A., Foster, E., Goldstein, S.A., Kuznetsova, T., Lancellotti, P., Muraru, D., Picard, M.H., Rietzschel, E.R., Rudski, L., Spencer, K.T., Tsang, W., Voigt, J.U.: Recommendations for cardiac chamber quantification by echocardiography in adults: An update from the American Society of Echocardiography and the European Association of Cardiovascular Imaging. European Heart Journal – Cardiovascular Imaging 16(3), 233–271 (2015) 7, 22 20. Li, H., Wang, Y., Qu, M., Cao, P., Feng, C., Yang, J.: Echoefnet: Multi-task deep learning network for automatic calculation of left ventricular ejection fraction in 2d echocardiography. Computers in Biology and Medicine 156, 106705 (2023) 2 21. Li, M., Lin, B., Chen, Z., Lin, H., Liang, X., Chang, X.: Auxiliary signal-guided knowledge encoder-decoder for medical report generation. World Wide Web 26(1), 253–270 (2023). https://doi.org/10.1007/s11280-022-01013-6 3 22. Lim, G.B., Leong, Y.Y., Tay, E.L., Chan, M.Y., Yeo, T.J., Lam, C.S., Januzzi, J.L., Richards, A.M., Jiang, B.: Global burden of cardiovascular diseases: Projections from 2025 to 2050. European Heart Journal (2024). https://doi.org/10.1093/ eurheartj/ehae602 2 23. Lin, C.Y.: ROUGE: A package for automatic evaluation of summaries. In: Text Summarization Branches Out. pp. 74–81 (2004) 8

EchoSonar-R

17

24. Liu, Z., et al.: Understanding r1-zero-like training: A critical perspective. arXiv preprint arXiv:2503.20783 (2025) 6 25. Maani, F., Ukaye, A., Saadi, N., Saeed, N., Yaqub, M.: Simlvseg: simplifying left ventricular segmentation in 2-d+ time echocardiograms with self-and weakly supervised learning. Ultrasound in Medicine & Biology 50(12), 1945–1954 (2024) 2 26. Maani, F.A., Saeed, N., Matsun, A., Yaqub, M.: Coreecho: Continuous representation learning for 2d+ time echocardiography analysis. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 591–601. Springer (2024) 2 27. Mitchell, C., Rahko, P.S., Blauwet, L.A., Canaday, B., Finstuen, J.A., Foster, M.C., Horton, K., Ogunyankin, K.O., Palma, R.A., Velazquez, E.J.: Guidelines for performing a comprehensive transthoracic echocardiographic examination in adults: Recommendations from the American Society of Echocardiography. Journal of the American Society of Echocardiography 32(1), 1–64 (2019). https: //doi.org/10.1016/j.echo.2018.06.004 2, 5 28. Muhtaseb, R., Yaqub, M.: Echocotr: Estimation of the left ventricular ejection fraction from spatiotemporal echocardiography. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 370–379. Springer (2022) 2 29. Ostmeier, S., Xu, J., Chen, Z., Varma, M., Blankemeier, L., Bluethgen, C., Michalson, A.E., Moseley, M., Langlotz, C., Chaudhari, A.S., Delbrouck, J.B.: GREEN: Generative radiology report evaluation and error notation. In: Findings of the Association for Computational Linguistics: EMNLP 2024. pp. 374–390 (2024). https://doi.org/10.18653/v1/2024.findings-emnlp.21 8, 11, 26 30. Ouyang, D., He, B., Ghorbani, A., Yuan, N., Ebinger, J., Langlotz, C.P., Heidenreich, P.A., Harrington, R.A., Liang, D.H., Ashley, E.A., Zou, J.Y.: Video-based AI for beat-to-beat assessment of cardiac function. Nature 580(7802), 252–256 (2020). https://doi.org/10.1038/s41586-020-2145-8 7, 9, 27 31. Ouyang, L., Wu, J., Jiang, X., Almeida, D., Wainwright, C., Mishkin, P., Zhang, C., Agarwal, S., Slama, K., Ray, A., et al.: Training language models to follow instructions with human feedback. Advances in Neural Information Processing Systems 35, 27730–27744 (2022) 5 32. Pan, J., Liu, C., Wu, J., Liu, F., Zhu, J., Li, H.B., Chen, C., Ouyang, C., Rueckert, D.: MedVLM-R1: Incentivizing medical reasoning capability of vision-language models via reinforcement learning. arXiv preprint arXiv:2502.19634 (2025) 3 33. Papineni, K., Roukos, S., Ward, T., Zhu, W.J.: BLEU: A method for automatic evaluation of machine translation. In: Proceedings of the 40th Annual Meeting of the Association for Computational Linguistics (ACL). pp. 311–318 (2002) 8 34. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., Sutskever, I.: Learning transferable visual models from natural language supervision. In: Proceedings of the 38th International Conference on Machine Learning (ICML). pp. 8748–8763 (2021) 3 35. Rahman, S., Haque, R., Swapno, S.M.R., Islam, M.B., Nobel, S.N., et al.: Deep learning-based left ventricular ejection fraction estimation from echocardiographic videos. In: 2023 International Conference on Evolutionary Algorithms and Soft Computing Techniques (EASCT). pp. 1–6. IEEE (2023) 2 36. Reale-Nosei, G., Amador-Domínguez, E., Serrano, E.: From vision to text: A comprehensive review of natural image captioning in medical diagnosis and radiology report generation. Medical Image Analysis 97, 103264 (2024). https: //doi.org/10.1016/j.media.2024.103264 8

18

Taratynova et al.

37. Roth, G.A., Mensah, G.A., Fuster, V.: The global burden of cardiovascular diseases and risks: A compass for global action. Journal of the American College of Cardiology 82(25), 2350–2473 (2023). https://doi.org/10.1016/j.jacc.2023.11.007 1 38. Salih, A., Boscolo Galazzo, I., Raisi-Estabragh, Z., Petersen, S.E., Menegaz, G., Salih, A.: Explainable artificial intelligence and cardiac imaging: Toward more interpretable models. Circulation: Cardiovascular Imaging 16(4), e014519 (2023). https://doi.org/10.1161/CIRCIMAGING.122.014519 2 39. Savarese, G., Becher, P.M., Lund, L.H., Seferovic, P., Rosano, G.M., Coats, A.J.: Cardiovascular health care in low- and middle-income countries. Circulation 149, e224–e248 (2024). https://doi.org/10.1161/CIRCULATIONAHA.123.065717 2 40. Sellergren, A., Kazemzadeh, S., Jaroensri, T., Kiraly, A., Traverse, M., Kohlberger, T., et al.: MedGemma technical report. arXiv preprint arXiv:2507.05201 (2025) 3, 9, 27 41. Shao, Z., Wang, P., Zhu, Q., Xu, R., Song, J., Zhang, M., Li, Y., Wu, Y., Guo, D.: DeepSeekMath: Pushing the limits of mathematical reasoning in open language models. arXiv preprint arXiv:2402.03300 (2024) 5, 6 42. She, C., Lu, R., Chen, L., Wang, W., Huang, Q.: EchoVLM: Dynamic mixture-ofexperts vision-language model for universal ultrasound intelligence. arXiv preprint arXiv:2509.14977 (2025) 3, 9, 27 43. Sun, H., et al.: Chiron-o1: Igniting multimodal large language models towards generalizable medical reasoning via mentor-intern collaborative search. In: Advances in Neural Information Processing Systems (NeurIPS) (2025) 3, 9, 27 44. Syryca, F., Gräßer, C., Trenkwalder, T., et al.: Automated generation of echocardiography reports using artificial intelligence: A novel approach to streamlining cardiovascular diagnostics. The International Journal of Cardiovascular Imaging 41, 967–977 (2025). https://doi.org/10.1007/s10554-025-03382-1 3 45. Team, G., Riviere, M., Pathak, S., Sessa, P.G., Hardin, C., Bhupatiraju, S., Hussenot, L., Mesnard, T., Shahriari, B., Ramé, A., et al.: Gemma 2: Improving open language models at a practical size. arXiv preprint arXiv:2408.00118 (2024) 22 46. Teed, Z., Deng, J.: RAFT: Recurrent all-pairs field transforms for optical flow. In: European Conference on Computer Vision (ECCV). pp. 402–419 (2020) 7, 20 47. Thapa, R., Li, A., Wu, Q., He, B., Sahashi, Y., Binder-Rodriguez, C., Zhang, A., Ouyang, D., Zou, J.: Mimic-iv-echo-ext-mimicechoqa: A benchmark dataset for echocardiogram-based visual question answering (2025) 7, 9, 27 48. Timmis, A., Vardas, P., Townsend, N., Torbica, A., Katus, H., De Smedt, D., Broccoli, S., Hinber, B., Ziegler, J., Maggioni, A.P., et al.: European society of cardiology: Cardiovascular disease statistics 2024. European Heart Journal 45(40), 4017–4184 (2024). https://doi.org/10.1093/eurheartj/ehae466 2 49. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, Ł., Polosukhin, I.: Attention is all you need. In: Advances in Neural Information Processing Systems (NeurIPS). vol. 30 (2017) 3 50. Vinyals, O., Toshev, A., Bengio, S., Erhan, D.: Show and tell: A neural image caption generator. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 3156–3164 (2015) 3 51. Vukadinovic, M., Xu, A., Cheng, X., Kwan, A.C., Ouyang, D.: EchoPrime: A multivideo view-informed vision-language model for comprehensive echocardiography interpretation. arXiv preprint arXiv:2410.09704 (2024) 2, 3, 4, 8

EchoSonar-R

19

52. Wei, J., Wang, X., Schuurmans, D., Bosma, M., Ichter, B., Xia, F., Chi, E., Le, Q.V., Zhou, D.: Chain-of-thought prompting elicits reasoning in large language models. In: Advances in Neural Information Processing Systems (NeurIPS). vol. 35 (2022) 3 53. Won, D., Walker, J., Horowitz, R., Bharadwaj, S., Carlton, E., Gabriel, H.: Sound the alarm: The sonographer shortage is echoing across healthcare. Journal of Ultrasound in Medicine 43(7), 1289–1301 (2024). https://doi.org/10.1002/jum. 16453 2 54. World Health Organization: Cardiovascular diseases (CVDs): Fact sheet. https: //www.who.int/news- room/fact- sheets/detail/cardiovascular- diseases(cvds) (2024), accessed: 2025-05-01 1 55. Xu, K., Ba, J., Kiros, R., Cho, K., Courville, A., Salakhutdinov, R., Zemel, R., Bengio, Y.: Show, attend and tell: Neural image caption generation with visual attention. In: Proceedings of the 32nd International Conference on Machine Learning (ICML). pp. 2048–2057 (2015) 3 56. Xu, W., Chan, H.P., Li, L., Aljunied, M., Yuan, R., Wang, J., Xiao, C., Chen, G., Liu, C., Li, Z., et al.: Lingshu: A generalist foundation model for unified multimodal medical understanding and reasoning. arXiv preprint arXiv:2506.07044 (2025) 3, 9, 27 57. Xu, Z., Wang, Z., Jiang, X., Li, D., Cheng, D., Wang, N.: Reasoning-driven multimodal LLM for domain generalization. arXiv preprint arXiv:2602.23777 (2026) 6 58. Yan, B., Liu, R., Kuo, D., Adithan, S., Reis, E., Kwak, S., Venugopal, V., O’Connell, C., Saenz, A., Rajpurkar, P., et al.: Style-aware radiology report generation with radgraph and few-shot prompting. In: Findings of the Association for Computational Linguistics: EMNLP 2023. pp. 14676–14688 (2023) 3 59. Yang, A., Yang, B., Zhang, B., Wang, B., et al.: Qwen3 technical report. arXiv preprint arXiv:2505.09388 (2025) 5, 8 60. Yu, Q., et al.: Dapo: An open-source llm reinforcement learning system that goes beyond. arXiv preprint arXiv:2503.14476 (2025) 6 61. Zhang, J., Gajjala, S., Agrawal, P., Tison, G.H., Hallock, L.A., Beussink-Nelson, L., Lassen, M.H., Fan, E., Aras, M.A., Jordan, C., et al.: Fully automated echocardiogram interpretation in clinical practice: feasibility and diagnostic accuracy. Circulation 138(16), 1623–1635 (2018) 2 62. Zhang, T., Kishore, V., Wu, F., Weinberger, K.Q., Artzi, Y.: BERTScore: Evaluating text generation with BERT. In: International Conference on Learning Representations (ICLR) (2020) 8 63. Zhao, Y., Lv, W., Xu, S., Wei, J., Wang, G., Dang, Q., Liu, Y., Chen, J.: DETRs beat YOLOs on real-time object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 16965– 16974 (2024) 4, 8 64. Zheng, L., Chiang, W.L., Sheng, Y., Zhuang, S., Wu, Z., Zhuang, Y., Lin, Z., Li, Z., Li, D., Xing, E.P., et al.: Judging LLM-as-a-judge with MT-Bench and Chatbot Arena. Advances in Neural Information Processing Systems (NeurIPS) 36 (2024) 8

20

Taratynova et al.

Table S1: Private dataset statistics across train and test splits. Left: abnormality label and view distributions. Right: QA pair distribution and detector annotation statistics before and after RAFT propagation [46]. † GRPO uses only Abnormality List and Full Report question types. Train (n = 5,061) Pos. Abnormality Labels TV Regurgitation MV Regurgitation AV Regurgitation LA Enlargement LV Systolic Dysfunction MV Calcification LV Enlargement AV Stenosis RA Enlargement Bicuspid AV RV Enlargement Normal

3,291 3,133 1,191 1,694 1,070 701 504 481 384 230 221 556

Neg. 1,770 1,928 3,870 3,367 3,991 4,360 4,557 4,580 4,677 4,831 4,840 –

Test (n = 1,215) Pos. 662 659 188 330 198 105 70 116 59 31 31 212

Neg. 553 556 1,027 885 1,017 1,110 1,145 1,099 1,156 1,184 1,184 –

View Distribution A2C A2C Color on MV A3C A3C Color on AV A4C A4C Zoomed LV PLAX Standard PLAX Mitral Cusps PSAX Zoomed Out PSAX Color on Pulm. Total

Clips

%

Clips

%

5,633 2,997 6,069 2,351 7,087 3,374 7,485 3,158 5,665 2,563 46,382

12.1 6.5 13.1 5.1 15.3 7.3 16.1 6.8 12.2 5.5 100

1,409 693 1,485 548 1,911 750 1,770 780 1,316 648 11,310

12.5 6.1 13.1 4.8 16.9 6.6 15.6 6.9 11.6 5.7 100

A

Supplementary Material

A.1

Private Dataset

Train Question Type QA Pairs Structure Description Abnormality Classification Abnormality List Conclusion Full Report Total

%

Pairs

%

57,622 55,671 5,061 5,061 4,800 128,215

44.9 43.4 3.9 3.9 3.7 100

14,230 13,365 1,215 1,215 1,184 31,209

45.6 42.8 3.9 3.9 3.8 100

Detector Annotations, n frames Pre-RAFT Left Ventricle (LV) Left Atrium (LA) Right Atrium (RA) Right Ventricle (RV) Mitral Valve (MV) Tricuspid Valve (TV) LVOT Total

Test

Pairs

Post-RAFT

Annotated

Train

Test

35,365 7,907 3,958 3,018 1,309 1,294 519 53,370

319,375 86,207 36,290 33,011 10,312 10,007 7,540 502,742

50,898 13,014 5,470 5,038 1,367 1,229 983 77,999

Pre-RAFT

Train

Test

10,291 384,456 1.5

7,964 332,047 1.52

1,245 52,674 1.48

Coverage Statistics DICOMs Total frames Avg. structs / frame

The private echocardiography dataset was collected between 2018 and 2022 from 139 sites and comprises 5,061 training and 1,215 test studies, containing 46,382 and 11,310 videos, respectively. Each study consists of multi-view echocardiographic clips acquired across up to 10 standard views. At training and inference time, one clip per unique view is randomly selected per study, providing implicit data augmentation across epochs. Per-study clip counts range from 4 to 30 in training and 4 to 22 at test time, with a median of 9 clips per study. Full dataset statistics are reported in Tab. S1. Abnormality Labels. Each study is annotated with binary labels for 12 cardiac abnormalities spanning valvular disease, chamber enlargement, systolic dysfunction, and structural anomalies. TV and MV Regurgitation are the most prevalent, while Bicuspid AV and RV Enlargement are the rarest. View Distribution. The dataset spans 10 standard echocardiographic views: seven B-mode views (PLAX Standard, A4C, A3C, PSAX Zoomed Out, A2C, A4C Zoomed LV, PLAX Mitral Cusps) and three color Doppler variants (A2C Color on MV, PSAX Color on Pulmonary, A3C Color on AV). The train and test splits follow closely matched distributions across all views.

EchoSonar-R

21

Table S2: Representative question–answer examples per type used in SFT and GRPO training. † Used in GRPO only. All examples are drawn from the private training set. Type

Example Question

Example Answer

Structure Description

What are the echocardiographic findings for the left atrium? Describe the tricuspid valve findings in this echocardiographic study. What does the echocardiogram reveal about the aortic root?

Normal dimensions. No suggestion of mass or thrombus.

Abnormality Classification

Is there tricuspid regurgitation present in this echocardiogram? Is the left atrium enlarged?

Normal appearance. Mild tricuspid regurgitation. Normal dimensions for BSA. Yes, tricuspid regurgitation is present. No, there is no evidence of left atrial enlargement.

Is there left ventricular systolic dysfunction present in this echocardiogram?

No, there is no evidence of left ventricular systolic dysfunction.

Abnormality List

Summarize the abnormalities observed in this echocardiographic examination. What abnormalities are present in this echocardiographic study?

The following abnormalities are identified: Tricuspid Regurgitation, Mitral Valve Regurgitation. No significant abnormalities identified.

Conclusion

What is the overall conclusion of this echocardiographic study? Summarize the key conclusions from this echocardiographic examination.

Normal left ventricular size and function. Mild mitral regurgitation. Stage I diastolic dysfunction with normal LV filling pressure. No evidence of hypertrophy, chamber enlargement, or significant valvular disease.

Full Report

Produce a detailed echocardiographic report with findings for each structure and a final conclusion.

Left Ventricle: Normal dimensions and systolic function. Normal wall thickness. Impaired LV relaxation consistent with Stage I diastolic dysfunction. LV filling pressure is normal. Right Ventricle: Normal dimensions and systolic function. Left Atrium: Normal dimensions. No suggestion of mass or thrombus. Right Atrium: Normal dimensions. No suggestion of mass or thrombus. Aortic Valve: Normal trileaflet morphology. Normal Doppler flow characteristics. Mitral Valve: Normal appearance without prolapse. Mild mitral regurgitation. Tricuspid Valve: Normal appearance. Normal Doppler flow characteristics. Pulmonic Valve: Normal appearance. Normal Doppler flow characteristics. Pericardium: Appears normal. No significant effusion is seen. Aortic Root: Normal dimensions for BSA. Aortic Arch: Normal dimensions. Normal Doppler flow profile. Pulmonary Artery: RVSP could not be estimated because of insufficient tricuspid regurgitation. Conclusions: Normal left ventricular size with estimated ejection fraction of 60%. No evidence of hypertrophy, chamber enlargement or aortic dilatation. Mild mitral regurgitation. RVSP could not be estimated because of insufficient tricuspid regurgitation. Stage I diastolic dysfunction with normal LV filling pressure. Normal peak velocities noted through the aortic valve, left ventricular outflow tract, pulmonic valve and descending aorta.

Training Data. For SFT, the training set yields 128,215 QA pairs across 5,061 studies (mean 25.3 pairs per study), spanning all five question types. For GRPO, only the abnormality list and full report question types are used, as these directly optimize for the two target tasks. Tab. S2 shows representative examples of each question type. Structure-Aware Detector Annotations. A subset of 2,689 training studies was manually annotated with segmentation masks for seven cardiac structures: LV, LA, RV, RA, MV, TV, and LVOT. Masks were converted to bounding boxes and propagated across all frames using RAFT optical flow, expanding from sparsely annotated keyframes to full per-frame coverage. A.2

Reasoning Trace Generation

Reasoning traces for all training samples are generated offline with Qwen3-8B in a text-only setting, with temperature = 0.6, top-p = 0.95. The model is given the correct answer upfront and asked to reason backward toward it.

22

Taratynova et al.

Prompt Construction. For each study, a structured prompt is constructed containing: (i) a system message identifying the model as an expert echocardiographer; (ii) the ground-truth disease label or list of present abnormalities; (iii) the list of available echocardiographic views; and for each view: (iv) detected cardiac structures, (v) pathology-specific assessment guidance, and (vi) clinical findings from the ground-truth report. The prompt concludes with an instruction to analyze each view and synthesize a final assessment. The assessment guidance is encoded as per-pathology diagnostic templates that specify, for each view, its diagnostic importance and the structures and features that are clinically relevant for a given condition. For example, the Aortic Stenosis template designates PLAX Standard for leaflet morphology and calcification assessment, PSAX Zoomed Out for valve orifice planimetry and cusp count, and A3C Color on AV for peak velocity and mean gradient via CW Doppler. Templates are defined for all 11 target abnormalities, 12 cardiac structures, and a comprehensive all-pathology view set, covering both B-mode and color Doppler views, and are derived from clinical guidelines. Reasoning Trace Construction. The generated analyses are mapped to VQA training samples as follows. For abnormality classification questions, the diseasespecific analysis for the queried condition is used. For structure description questions, the structure-specific section of the analysis is used. For abnormality list, conclusion, and full report questions, the analysis is generated using a viewcomplete template covering all cardiac structures across all views simultaneously. Each reasoning trace is wrapped in <think>. . .</think> tags and prepended to the ground-truth answer, forming the complete supervised target Y ∗ . Reasoning Trace Quality Audit. To validate the quality of the generated training traces, we audit 1,000 randomly sampled traces using two independent judges (Gemma-2-9B-it [45] and Phi-4-14B [1]) across six dimensions: fabrication, coverage, visual grounding, view-pathology appropriateness, internal consistency, and terminology accuracy. Both judges score visual grounding and view-pathology appropriateness below terminology accuracy (Gemma: 4.885, 4.860 vs. 4.998; Phi-4: 4.872, 4.942 vs. 4.980), indicating the audit is not at ceiling; all dimensions nonetheless score above 4.73/5. These results confirm that the LLM structures clinical content already present in the prompt rather than fabricating new findings, as the traces are anchored on per-pathology assessment templates derived from American Society of Echocardiography guidelines [19], ground-truth report findings, detector-confirmed structures, and the available view list. A.3

Structure-Aware Detector

Architecture and Training Data. The structure-aware detector is an RT-DETR-L model fine-tuned from COCO-pretrained weights to localize seven cardiac structures: LV, LA, RA, RV, MV, TV, and LVOT. Training uses RAFT-propagated bounding boxes ( Sec. A.1) at 640 × 640 resolution. Given the substantial annotation imbalance between common structures (LV, LA) and rarer ones (RA, RV,

EchoSonar-R

23

MV, TV, LVOT), we apply class-balanced sampling: all frames containing rare structures are retained, with an equal number of LV/LA-only frames randomly sampled to match. Training Configuration. The detector is trained for 5 epochs with batch size 64, AdamW optimizer (lr0 = 0.01, lrf = 0.01), and AMP enabled. Augmentation includes mosaic (p = 1.0), horizontal flip (p = 0.5), random scale (±0.5), random translate (±0.1), random erasing (p = 0.4), HSV jitter, and RandAugment. IoU threshold is set to 0.7. Best validation performance is achieved at epoch 2, with mAP50 = 0.350 and mAP50-95 = 0.161 (precision 0.304, recall 0.619). Qualitative Results. Fig. S1 shows representative detection results on the private dataset. The model correctly localises the dominant structures (LV, RV, LA) across views, with predicted boxes closely matching ground truth. Smaller structures such as MV and TV are detected with lower confidence but remain anatomically consistent. Fig. S2 demonstrates generalisation to MIMICEchoQA and EchoNet-Dynamic, both unseen during training. Despite differences in image quality, scanner settings, and the presence of color Doppler overlays, the detector produces anatomically plausible localisations across all four examples, suggesting that the learned structural representations transfer across institutions and acquisition protocols. A.4

View Classifier

The view classifier assigns a standard echocardiographic view label to each incoming video, providing the textual view identifiers interleaved with visual tokens in the multi-view input sequence (Sec. 3.1). It is based on MobileNet-V3-Small, fine-tuned from ImageNet-pretrained weights on the private training set using a single random frame per video per epoch. Training uses AdamW with differential learning rates (backbone 10−4 , head 10−3 ) and cosine annealing over 15 epochs. On the private test set, the classifier achieves 95.7% accuracy and 95.7% macro F1, with well-calibrated confidence (ECE = 0.0078). Per-view F1 ranges from 0.888 (A4C Zoomed LV) to 0.992 (PSAX Color on Pulmonary). The most frequent confusions occur between visually similar views: A4C and A4C Zoomed LV, A2C and A4C, which share similar apical geometry and are distinguishable primarily by field of view. A.5

Evaluation Prompts

At evaluation time, all models are prompted to perform two tasks: abnormality classification and report generation. Prompts differ between EchoSonar-R and the baseline VLMs to accommodate their respective input formats and output conventions. Abnormality Classification. EchoSonar-R and its ablations are prompted with an open-ended question, consistent with the SFT training format:

24

Taratynova et al.

Fig. S1: Qualitative detection results on the private dataset. In each pair, the left image shows ground-truth bounding boxes and the right shows model predictions with confidence scores. Colour coding: ■ LV, ■ LA, ■ RA, ■ RV, ■ MV, ■ TV, ■ LVOT.

What abnormalities are present in this echocardiographic study? Baseline VLMs receive a structured prompt to target abnormalities explicitly, as these models have not been trained to produce a free-form abnormality list: Based on this echocardiographic study, identify which of the following abnormalities are present or absent: Tricuspid Regurgitation, Mitral Valve Regurgitation, Aortic Stenosis, Mitral Valve Calcification, Left Atrial Enlargement, Left Ventricular Enlargement, Left Ventricular Systolic Dysfunction, Right Atrial Enlargement, Right Ventricular Enlargement, Bicuspid Aortic Valve, Aortic Regurgitation For each abnormality, state whether it is Present or Absent. Report Generation. For report generation, the following prompt is used:

EchoSonar-R

25

MIMICEchoQA

EchoNet-Dynamic Fig. S2: Zero-shot generalisation of the structure-aware detector to public benchmarks unseen during training. Colour coding as in Fig. S1.

You are an expert echocardiographer. Based on the provided echocardiographic images, generate a structured echocardiographic report with the following sections. For each section, provide a concise clinical description of the findings. Left Ventricle: Right Ventricle: Left Atrium: Right Atrium: Aortic Valve: Mitral Valve: Tricuspid Valve: Pulmonic Valve: Pericardium: Aortic Root: Aortic Arch: Pulmonary Artery: Conclusions: Write each section as a brief clinical narrative. If a structure is not well visualized, state that.

26

Taratynova et al.

Predicted abnormality lists and report sections are extracted from model outputs via string matching and section parsing, respectively. For EchoSonar-R, only the content after </think> is used for evaluation; the reasoning trace is excluded. Reasoning Quality Evaluation. Reasoning trace quality is assessed using an LLM-as-a-judge framework on five clinical dimensions, each rated on a 1–5 Likert scale. The evaluation prompt provides the original question, ground-truth answer, and full model response. The five dimensions are: 1. Reasoning Efficiency: whether every reasoning step advances toward the answer, penalising redundancy and circular reasoning. 2. Factual Correctness: accuracy of clinical and anatomical claims made during reasoning. 3. Evidence Grounding: whether the model references specific visual observations rather than relying on generic statements. 4. Terminology Accuracy: correct use of echocardiographic and medical terminology throughout. 5. Reasoning-Answer Agreement: consistency between findings discussed in the reasoning trace and those stated in the final answer, penalising both omissions and unsupported additions. Clinical Report Faithfulness. Report generation quality is assessed using an echocardiography-adapted version of the GREEN metric [29], evaluated sectionby-section. For each predicted report section, the judge identifies clinically significant errors (false findings, omissions, wrong anatomical location, misassessed severity, incorrect comparison to prior studies) and clinically insignificant errors (omission of minor findings, wording differences that do not change clinical meaning), and counts matched findings. The GREEN score is then computed from these counts as described in [29]. This approach is more sensitive to clinical correctness than lexical metrics such as BLEU or ROUGE, which penalise semantically equivalent phrasings. The full evaluation prompt is shown below. A.6

Additional Results

Per-Abnormality Results. Tab. S3 reports per-abnormality F1, balanced accuracy (BAcc), sensitivity, and specificity across all three evaluation sets. On the private test set, EchoSonar-R achieves the highest F1 on 9 of 12 conditions, with the largest margins on rare or structurally complex abnormalities such as Aortic Stenosis, Bicuspid AV, and LV Enlargement. The high specificity across these rare conditions (98-99%) confirms that the model avoids over-predicting lowprevalence pathologies. On MIMICEchoQA, EchoSonar-R leads on LV Systolic Dysfunction and Aortic Regurgitation, though performance on Right Ventricular Enlargement and Left Atrial Enlargement is lower, consistent with the label shift discussed in the main paper. On EchoNet-Dynamic, which evaluates only EFbased LV systolic dysfunction from a single apical view, EchoSonar-R achieves the highest BAcc (60.2%).

EchoSonar-R

27

Table S3: Abnormality classification performance across private and public datasets. We report per-disease F1 score (↑), balanced accuracy (BAcc, ↑), sensitivity (Se, ↑), and specificity (Sp, ↑). EchoSonar-R† denotes GRPO training; EchoSonar-R∗ denotes SFT-only. Baselines: Qwen3-VL [5], MedGemma [40], EchoVLM [42], Chiron-o1 [43], Lingshu [56]. Prev. = prevalence (%). EchoSonar-R† Abnormality

F1

Prev.

BAcc

Se

EchoSonar-R∗

Qwen3-VL

MedGemma

EchoVLM

Chiron-o1

Lingshu

Sp

F1

BAcc

Se

Sp

F1

BAcc

Se

Sp

F1

BAcc

Se

Sp

F1

BAcc

Se

Sp

F1

BAcc

Se

Sp

F1

BAcc

Se

Sp

59.9 61.6 75.8 62.3 74.0 65.4 82.2 72.2 75.5 58.7 55.5 65.7 67.4

71.5 74.3 64.8 35.2 52.8 39.4 66.4 51.4 52.9 18.6 12.9 32.3 47.7

48.3 48.9 86.9 89.4 95.2 91.5 98.0 92.9 98.1 98.7 98.1 99.2 87.1

66.6 69.1 62.7 36.2 55.9 41.3 69.8 44.5 53.9 18.4 3.6 19.2 45.1

59.3 62.6 74.6 61.3 71.9 65.1 80.9 69.7 71.5 55.9 50.6 57.4 65.1

72.4 75.1 65.5 33.0 49.0 39.9 63.8 44.8 44.3 13.6 3.2 16.1 43.4

46.3 50.0 83.8 89.5 94.9 90.3 98.0 94.7 98.8 98.3 98.0 98.6 86.8

38.3 49.7 26.6 20.1 21.8 24.2 14.9 15.4 11.2 6.0 2.5 4.5 19.6

52.1 51.5 49.8 49.9 51.8 52.3 50.3 51.7 52.5 46.9 44.2 50.1 50.3

28.5 44.6 26.4 24.1 26.8 41.5 31.9 45.7 42.9 16.9 12.9 19.4 30.1

75.6 58.5 73.3 75.7 76.8 63.2 68.7 57.7 62.1 76.9 75.6 80.8 70.4

27.9 24.6 35.1 23.7 19.6 16.8 10.0 9.4 8.8 9.1 2.3 6.7 16.2

51.2 49.2 49.3 48.1 48.6 50.4 49.7 48.1 47.9 50.9 41.1 55.6 49.2

18.4 16.1 51.2 43.9 28.3 17.6 11.2 14.3 31.4 50.8 16.1 35.5 27.9

83.9 82.4 47.5 52.2 68.8 83.2 88.2 82.0 64.5 50.9 66.1 75.8 70.5

0.0 0.0 0.0 0.0 0.0 26.8 5.8 0.0 0.0 0.0 0.0 0.0 2.7

50.0 50.0 50.0 50.0 50.0 50.0 50.9 50.0 50.0 50.0 50.0 50.0 50.1

0.0 0.0 0.0 0.0 0.0 100.0 3.4 0.0 0.0 0.0 0.0 0.0 8.6

100.0 100.0 100.0 100.0 100.0 0.0 98.3 100.0 100.0 100.0 100.0 100.0 91.5

61.1 69.6 40.2 0.0 27.9 26.1 16.6 14.1 11.1 8.9 5.4 5.2 23.9

48.8 50.5 49.5 49.9 50.5 50.0 50.9 47.7 51.1 48.4 53.8 51.9 50.3

70.8 96.1 79.7 0.0 92.4 82.4 51.7 57.1 92.9 88.1 100.0 67.7 73.2

26.8 5.0 19.2 99.8 8.6 17.5 50.1 38.2 9.4 8.7 7.7 36.1 27.3

62.3 33.6 29.5 1.9 28.0 25.6 18.5 15.9 5.0 8.7 5.2 5.1 19.9

49.4 50.5 48.0 50.4 51.2 49.9 54.7 51.2 46.2 48.0 52.1 51.4 50.3

73.4 24.1 35.8 0.9 86.4 75.0 30.2 76.2 10.0 76.3 83.9 67.7 53.3

25.3 76.8 60.3 99.8 16.1 24.7 79.3 26.2 82.4 19.6 20.3 35.0 47.2

52.7 64.0 60.1 60.6 57.5 59.0

28.6 33.3 29.4 30.0 18.2 27.9

76.9 94.7 90.8 91.1 96.9 90.1

6.1 31.1 22.7 8.3 31.6 20.0

45.8 60.3 59.7 52.1 62.5 56.1

14.3 25.9 29.4 10.0 27.3 21.4

77.4 94.7 89.9 94.2 97.8 90.8

58.3 53.3 30.0 52.6 27.6 44.4

37.5 49.1 58.8 46.4 51.1 48.6

50.0 25.0 66.7 44.4 53.8 0.0 17.6 100.0 30.0 50.0 42.9 18.2 36.4 65.9 0.0 39.7 57.5 23.0

32.1 50.0 58.8 55.0 50.0 49.2

64.3 0.0 17.6 10.0 0.0 18.4

0.0 0.0 100.0 80.6 100.0 0.0 100.0 0.0 100.0 0.0 80.0 16.1

50.0 50.0 50.0 50.0 47.6 49.5

0.0 100.0 66.7 100.0 0.0 65.5 0.0 100.0 78.9 0.0 100.0 72.0 0.0 95.1 27.0 20.0 79.0 62.0

53.6 44.9 66.8 52.1 47.1 52.9

57.1 66.7 88.2 90.0 45.5 69.5

50.0 23.1 45.5 14.3 48.8 36.3

86.7 38.1 62.9 44.4 28.6 52.1

58.9 37.9 50.5 41.4 55.1 48.8

92.9 29.6 64.7 40.0 27.3 50.9

25.0 46.2 36.4 42.9 82.9 46.7

31.1 89.1 37.2

58.7

27.1 90.3

5.6

50.6

3.0

50.5

98.4

2.5

50.0

100.0

51.8

35.0

68.6

7.7

50.0

4.3

95.6

Private Test Set (n = 1,215 studies, multi-view) TV Regurgitation MV Regurgitation LA Enlargement Healthy LV Systolic Dysfunction AV Regurgitation AV Stenosis MV Calcification LV Enlargement RA Enlargement RV Enlargement Bicuspid AV Macro Average

54.6 54.4 27.3 17.5 16.3 15.4 9.5 8.6 5.7 4.8 2.5 2.5

66.6 68.3 64.9 38.0 59.5 42.3 71.6 45.4 57.4 25.9 13.8 39.2 49.4

MIMICEchoQA [47] (single-view) LA Enlargement, n = 18 AV Regurgitation, n = 40 LV Syst. Dysfunc., n = 28 RV Enlargement, n = 17 AV Stenosis, n = 52 Macro Average

77.8 67.5 60.7 58.8 21.1

11.6 38.3 23.8 18.2 20.0 22.4

EchoNet-Dynamic [30] (n = 1,277, single-view) LV Systolic Dysfunction

39.3

60.2

98.2

51.2

66.7

0.0

36.0

Table S4: Echo-GREEN scores per report section on the private test set. Higher is better. EchoSonar-R† denotes GRPO training; EchoSonar-R∗ denotes SFT-only. Best score per section in bold. Qwen3-VL

MedGemma

Chiron-o1

Lingshu

EchoSonar-R∗

EchoSonar-R†

Left Ventricle Right Ventricle Left Atrium Right Atrium Aortic Valve Mitral Valve Tricuspid Valve Pulmonic Valve Pericardium Aortic Root Aortic Arch Pulmonary Artery Conclusions

0.394 0.181 0.259 0.255 0.195 0.210 0.086 0.078 0.371 0.354 0.149 0.049 0.226

0.542 0.760 0.331 0.351 0.117 0.308 0.226 0.392 0.961 0.920 0.500 0.172 0.311

0.427 0.527 0.399 0.444 0.263 0.256 0.245 0.329 0.493 0.550 0.341 0.120 0.263

0.541 0.547 0.551 0.706 0.403 0.220 0.276 0.443 0.918 0.754 0.469 0.202 0.354

0.758 0.953 0.789 0.969 0.811 0.676 0.646 0.796 0.988 0.937 0.951 0.494 0.578

0.758 0.959 0.795 0.965 0.817 0.672 0.658 0.804 0.987 0.936 0.944 0.527 0.578

Average

0.216

0.457

0.359

0.491

0.796

0.800

Section

Per-Section Report Generation. Tab. S4 reports Echo-GREEN scores broken down by report section. EchoSonar-R leads on all 13 sections, with the largest margins over baselines on valvular sections and Pulmonary Artery. Structurally simple sections such as Pericardium and Aortic Root are well-handled by most models. The Conclusions section remains the most challenging across all models, requiring integration findings across all cardiac structures. Qualitative Analysis. Fig. S3-Fig. S5 show representative EchoSonar-R predictions with full reasoning traces across three characteristic cases. Fig. S3 illustrates a label-shift false positive on MIMICEchoQA: the model identifies trace mitral and tricuspid regurgitation, but only tricuspid regurgitation carries a positive label under the annotation convention.

28

Taratynova et al.

Table S5: Per-class ablation results on the private test set. VT = video tokens; ST = structure tokens. w/ reasoning

w/o reasoning

Remove Both

Remove VT

Abnormality

F1

BAcc Sens Spec

F1

BAcc Sens Spec

F1

BAcc Sens

Spec

TV Regurgitation MV Regurgitation Aortic Stenosis MV Calcification RA Enlargement RV Enlargement LA Enlargement LV Systolic Dysfunction LV Enlargement Bicuspid AV AV Regurgitation Healthy Macro Average

66.6 69.1 69.8 44.5 18.4 3.6 62.7 55.9 53.9 19.2 41.3 36.2 45.1

59.3 62.6 80.9 69.7 55.9 50.6 74.6 71.9 71.5 57.4 65.1 61.3 65.1

65.6 70.4 59.4 48.1 12.3 11.4 56.6 54.4 43.1 31.6 41.0 38.5 44.4

55.1 62.8 77.3 79.8 53.5 53.1 70.1 70.5 65.3 68.0 67.5 63.3 65.5

58.8 57.2 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 10.2

49.5 48.2 50.0 50.0 50.0 50.0 50.0 49.7 50.0 50.0 50.0 50.0 50.2

34.5 55.2 34.1 56.9 100.0 0.0 100.0 0.0 100.0 0.0 100.0 0.0 100.0 0.0 99.4 0.0 100.0 0.0 100.0 0.0 100.0 0.0 100.0 0.0 90.2 10.5

72.4 75.1 63.8 44.8 13.6 3.2 65.5 49.0 44.3 16.1 39.9 33.0 43.4

46.3 50.0 98.0 94.7 98.3 98.0 83.8 94.9 98.8 98.6 90.3 89.5 86.8

75.3 78.6 58.6 71.4 8.5 6.5 51.2 44.9 31.4 38.7 58.0 48.1 47.6

34.9 47.0 95.9 88.1 98.5 99.8 88.9 96.1 99.1 97.2 77.1 78.4 83.4

64.4 62.4 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 10.3

F1

Remove ST

BAcc Sens

Spec

F1

BAcc Sens Spec

53.2 49.5 50.0 50.0 50.0 50.0 50.0 49.9 50.0 50.0 50.0 50.0 49.8

53.5 38.6 100.0 100.0 100.0 100.0 100.0 99.7 100.0 100.0 100.0 100.0 88.0

65.3 65.9 66.0 44.1 15.8 15.4 65.9 53.7 53.3 18.2 39.1 24.1 43.9

54.7 57.9 78.6 69.6 54.6 55.7 76.7 70.9 72.1 56.1 63.6 55.7 63.9

53.0 60.4 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 11.5

75.0 72.8 59.5 44.8 10.2 12.9 67.0 47.5 45.7 12.9 36.2 17.9 41.9

34.4 42.9 97.8 94.5 99.0 98.6 86.4 94.3 98.4 99.2 91.0 93.5 85.8

Fig. S4 shows a correct true-negative prediction on a fully normal private study. The model reasons systematically across six views, assessing all detected structures per view and arriving at consistent all-absent conclusions. Fig. S5 shows a largely correct true-positive case with one false positive. The model correctly identifies LV Systolic Dysfunction, MV Regurgitation, and TV Regurgitation, all supported by consistent cross-view evidence of borderline LV dilation with mild global hypokinesis. The false positive for LV Enlargement is attributable to the same “borderline dilated” characterisation. Tab. S5 extends the aggregate ablation results with per-class F1, balanced accuracy, sensitivity, and specificity across all five configurations. Removing reasoning targets increases sensitivity at the cost of specificity (macro Sens: 47.6% vs. 43.4%; macro Spec: 83.4% vs. 86.8%), confirming that reasoning supervision induces a more conservative decision rule. The effect is most pronounced for MV Calcification, where removing reasoning substantially increases sensitivity (71.4% vs. 44.8%) but at the cost of specificity (88.1% vs. 94.7%). Removing video tokens (Remove VT) collapses performance to near-random BAcc across all classes, with F1 dropping to zero for 10 of 12 conditions. The model defaults to predicting the negative class, as reflected by near-100% specificity across all conditions. Structure tokens (Remove ST) provide a complementary contribution: the detector improves F1 on 10 of 12 classes (mean ∆F1 = +2.9% across improved classes) and BAcc on 9 of 12 (mean ∆BAcc = +2.5%), with performance drops concentrated in diseases with the sparsest detector training data (RA Enlargement, Bicuspid AV).

EchoSonar-R

29

Reasoning Quality Prompt You are an expert echocardiographer evaluating the quality of clinical reasoning in echocardiographic analysis. Task given to the model: {question} Ground truth answer: {ground_truth} Model’s response: {response} Evaluate the response on these 5 dimensions (score 1-5 each): [Reasoning Efficiency] 5: Concise and focused - every step advances toward the answer 4: Mostly focused with minor digressions 3: Some unnecessary repetition or tangents 2: Significant redundancy or circular reasoning 1: Rambling, mostly irrelevant [Factual Correctness] 5: All clinical/anatomical facts are correct 4: One minor factual inaccuracy 3: A few factual errors that don’t affect the conclusion 2: Multiple factual errors, some affecting the conclusion 1: Pervasive factual errors [Evidence Grounding] 5: Consistently references specific observations (chamber sizes, wall motion, valve appearance, Doppler patterns) 4: References some specific observations 3: Mix of specific and generic statements 2: Mostly generic reasoning with few image references 1: No reference to actual image content [Terminology Accuracy] 5: All medical/echocardiographic terms used correctly 4: One minor terminological imprecision 3: A few terminology issues that don’t cause confusion 2: Multiple misused terms (e.g., stenosis vs. regurgitation) 1: Pervasive terminology errors [Reasoning-Answer Agreement] 5: Perfect agreement - every finding in reasoning reflected in answer 4: Minor gap - one finding missing from answer or lacking reasoning 3: Partial agreement - some findings dropped or answer includes items not discussed in reasoning 2: Poor agreement - reasoning and answer substantially diverge 1: No agreement - reasoning and answer contradict each other Respond in this format: [Reasoning Efficiency] 〈score〉 [Factual Correctness] 〈score〉 [Evidence Grounding] 〈score〉 [Terminology Accuracy] 〈score〉 [Reasoning-Answer Agreement] 〈score〉 [Explanation] 〈brief explanation〉

30

Taratynova et al.

Echo-GREEN Prompt You are an expert echocardiographer evaluating the accuracy of a candidate echocardiographic report section against a reference. Your task: 1. Identify all clinically significant errors in the candidate. 2. Identify all clinically insignificant errors in the candidate. 3. Count the number of matched findings (correct statements). [Clinically Significant Errors] (a) False finding in candidate not present in reference. (b) Missing finding present in reference but absent from candidate. (c) Wrong anatomical location (e.g., wrong valve or chamber). (d) Misassessed severity or function (e.g., mild vs. severe). (e) False comparison to a prior study not in the reference. (f) Omitted comparison to a prior study mentioned in the reference. [Clinically Insignificant Errors] (a) Omission of a clinically insignificant finding. (b) Inclusion of a clinically insignificant finding not in reference. (c) Minor wording differences that do not change clinical meaning. Respond in this format: [Clinically Significant Errors] (a) 〈count〉 · · · (f) 〈count〉 Total: 〈total〉 [Clinically Insignificant Errors] (a) 〈count〉 · · · (c) 〈count〉 Total: 〈total〉 [Matched Findings] 〈count〉 [Explanation] 〈brief explanation〉 Reference: {reference} Candidate: {candidate}

EchoSonar-R

31

MIMICEchoQA Available view: A2C

Detected: LA, LV, RA, RV

Reasoning trace: View: A2C | Detected Structures: LA, LV, RA, RV Clinical Findings: Left Ventricle (LV): Normal dimensions and systolic function. Normal wall thickness. Diastolic filling pattern is normal. Right Ventricle (RV): Normal dimensions and systolic function. Left Atrium (LA): Normal dimensions. No suggestion of mass or thrombus. Right Atrium (RA): Normal dimensions. No suggestion of mass or thrombus. Aortic Valve (AV): Normal trileaflet morphology. Normal Doppler flow characteristics. Mitral Valve (MV): Normal appearance without prolapse. Trace mitral regurgitation. Tricuspid Valve (TV): Normal appearance. Trace tricuspid regurgitation. Pulmonic Valve (PV): Normal appearance. Normal Doppler flow characteristics. Pericardium: Appears normal. No significant effusion is seen. Aortic Root: Normal dimensions for BSA. Aortic Arch: Normal dimensions. Normal Doppler flow profile. Pulmonary Artery: Pulmonary artery systolic pressure could not be estimated due to insufficient tricuspid regurgitation. Interpretation: The A2C view provides a comprehensive assessment of the left and right ventricles, atria, and the aortic and tricuspid valves. The normal dimensions and systolic function of the LV and RV, along with normal wall thickness and diastolic filling patterns, suggest no significant structural or functional abnormalities. The normal appearance of the aortic and pulmonary valves indicates no significant valvular disease. The presence of trace mitral and tricuspid regurgitation is noted, which is often seen in the absence of significant structural abnormalities and may be due to minor valve leaflet incompetence or minor annular dilation. The normal dimensions and absence of mass or thrombus in the atria suggest no significant atrial pathology. The inability to estimate pulmonary artery systolic pressure due to insufficient tricuspid regurgitation suggests that the tricuspid regurgitation is minimal and does not significantly affect pulmonary artery pressure. Conclusions: Normal LV and RV dimensions and systolic function. Trace mitral and tricuspid regurgitation noted, likely due to minor valve leaflet incompetence. Normal atrial dimensions. Normal aortic root and ascending aorta. No pericardial effusion.

Predicted: TV Regurgitation, MV Regurgitation Ground truth: TV Regurgitation

Fig. S3: Qualitative example from MIMICEchoQA.

32

Taratynova et al.

Private set Available views: PLAX Standard, PSAX Zoomed Out, A4C, A4C Zoomed LV, A2C, A3C Reasoning trace: View: PLAX Standard | Detected: LA, LV, LVOT LV: No significant abnormalities in internal dimensions, septal and posterior wall thickness, global and regional systolic function, LVOT diameter, or EPSS. LA: No significant enlargement or structural abnormalities. AV: No significant calcification, thickening, or leaflet excursion issues. MV: No significant calcification, subvalvular apparatus issues, or annular diameter abnormalities. Pericardium: No effusion detected. View: PSAX Zoomed Out | Detected: LA, LV, LVOT, RA, RV LV: No significant abnormalities in short-axis wall motion, fractional shortening, regional wall motion abnormalities, or septal flattening. AV: No significant abnormalities in valve orifice area, number of cusps, or commissural fusion. TV: No significant tricuspid regurgitation noted. View: A4C | Detected: LA, LV, MV, RA, RV, TV LV: No significant abnormalities in endocardial tracing for volumes, LVEF calculation, global longitudinal strain, or apical wall motion. MV: No significant abnormalities in leaflet morphology or mitral inflow assessment. TV: No significant tricuspid regurgitation noted. View: A4C Zoomed LV | Detected: LV LV: No significant abnormalities in LVEF calculation, global longitudinal strain, or apical wall motion. View: A2C | Detected: LA, LV LV: No significant abnormalities in anterior and inferior wall motion, endocardial tracing for biplane volumes, or apical contractility. View: A3C | Detected: LV LV: No significant abnormalities in anteroseptal and inferolateral wall motion, longitudinal contractility, or LVOT morphology. AV: No significant abnormalities in leaflet separation and Doppler alignment. Conclusions: No significant abnormalities in any cardiac structures or functions assessed. LV shows normal dimensions, wall thickness, and systolic function. LA is of normal size. AV and MV show no significant pathology. RV and RA are of normal size and function with no significant tricuspid regurgitation. No pericardial effusion. Overall, the echocardiogram is normal.

Predicted: Healthy Ground truth: Healthy

Fig. S4: Qualitative example from the private test set.

EchoSonar-R

33

Private set Available views: PLAX Standard, PSAX Zoomed Out, A2C, A3C

Question: What abnor-

malities are present? Reasoning trace: View: PLAX Standard | Detected: LV, LA, AV, AR, MV, RV, Pericardium LV: Borderline dilated (taking patient size into account) with mild global hypokinesis. Possible anterior septal hypokinesis. RV: Normal dimensions and systolic function. LA: Normal dimensions. No suggestion of mass or thrombus. RA: Normal dimensions. No suggestion of mass or thrombus. AV: Normal trileaflet morphology. Normal Doppler flow characteristics. MV: Normal appearance without prolapse. Mild mitral regurgitation. TV: Normal appearance. Mild-to-moderate tricuspid regurgitation. Pericardium: Appears normal. No significant effusion. Aortic Root: Normal dimensions for BSA. Pulmonary Artery: Systolic pressure estimated to be normal. Implications: The borderline dilated LV with mild global hypokinesis and possible anterior septal hypokinesis suggests early systolic dysfunction, possibly related to coronary artery disease or hypertension. The mild MV and TV regurgitation indicate mild valvular insufficiency, possibly secondary to LV dysfunction. View: PSAX Zoomed Out | Detected: LV, LA, AV, MV, RV, TV, RA, Pericardium LV: Borderline dilated with mild global hypokinesis. Possible anterior septal hypokinesis. Findings consistent with PLAX view, reinforcing the presence of mild global hypokinesis. AV: Normal trileaflet morphology. Normal Doppler flow characteristics. MV: Normal appearance without prolapse. Mild mitral regurgitation. TV: Normal appearance. Mild-to-moderate tricuspid regurgitation. Pericardium: No significant effusion. View: A2C | Detected: LV, LA, MV LV: Borderline dilated with mild global hypokinesis. Possible anterior septal hypokinesis. Findings consistent with the PLAX and PSAX views, reinforcing mild global hypokinesis. LA: Normal dimensions. No suggestion of mass or thrombus. MV: Normal appearance without prolapse. Mild mitral regurgitation. View: A3C | Detected: LV, AV, LA, AA, MV, LVOT LV: Borderline dilated with mild global hypokinesis. Possible anterior septal hypokinesis. Anteroseptal and inferolateral wall motion consistent with prior views. AV: No significant abnormalities in leaflet separation or Doppler alignment. MV: Normal coaptation point. No mitral-aortic curtain calcification.

Predicted: TV Regurgitation, MV Regurgitation, LV Systolic Dysfunction, LV Enlargement Ground truth: TV Regurgitation, MV Regurgitation, LV Systolic Dysfunction Fig. S5: Qualitative example from the private test set.