Skip to main content An official website of the United States government Here's how you know Here's how you know Official websites use .gov A .gov website belongs to an official government organization in the United States. Secure .gov websites use HTTPS A lock ( Lock Locked padlock icon ) or https:// means you've safely connected to the .gov website. Share sensitive information only on official, secure websites. Search Log in Dashboard Publications Account settings Log out Search… Search NCBI Primary site navigation Search Logged in as: Dashboard Publications Account settings Log in Search PMC Full-Text Archive Search in PMC Journal List User Guide PERMALINK Copy As a library, NLM provides access to scientific literature. Inclusion in an NLM database does not imply endorsement of, or agreement with, the contents by NLM or the National Institutes of Health. Learn more: PMC Disclaimer | PMC Copyright Notice PLOS Digit Health . 2026 Apr 8;5(4):e0001309. doi: 10.1371/journal.pdig.0001309 Search in PMC Search in PubMed View in NLM Catalog Add to search Segment anything small for ultrasound: Enhancing segmentation with non-generative augmentation Danielle L Ferreira Danielle L Ferreira 1 Science and Technology Office, GE HealthCare, San Ramon, California, United States of America Conceptualization, Data curation, Formal analysis, Investigation, Methodology, Project administration, Software, Validation, Visualization, Writing – original draft, Writing – review & editing Find articles by Danielle L Ferreira 1, * , Ahana Gangopadhyay Ahana Gangopadhyay 1 Science and Technology Office, GE HealthCare, San Ramon, California, United States of America Conceptualization, Data curation, Software, Writing – review & editing Find articles by Ahana Gangopadhyay 1 , Hsi-Ming Chang Hsi-Ming Chang 1 Science and Technology Office, GE HealthCare, San Ramon, California, United States of America Project administration, Writing – review & editing Find articles by Hsi-Ming Chang 1 , Ravi Soni Ravi Soni 1 Science and Technology Office, GE HealthCare, San Ramon, California, United States of America Conceptualization, Funding acquisition, Investigation, Supervision, Writing – review & editing Find articles by Ravi Soni 1 , Gopal Avinash Gopal Avinash 1 Science and Technology Office, GE HealthCare, San Ramon, California, United States of America Funding acquisition, Writing – review & editing Find articles by Gopal Avinash 1 Editor: Ismini Lourentzou 2 Author information Article notes Copyright and License information 1 Science and Technology Office, GE HealthCare, San Ramon, California, United States of America 2 University of Illinois Urbana-Champaign, UNITED STATES OF AMERICA I have read the journal’s policy, and the authors of this manuscript have the following competing interests: all authors were employed by GE Healthcare during the research and manuscript preparation. However, the research was conducted independently, and the authors declare that the results and conclusions are presented objectively, without external influence from their employer. ✉ * E-mail: [email protected] Roles Danielle L Ferreira : Conceptualization, Data curation, Formal analysis, Investigation, Methodology, Project administration, Software, Validation, Visualization, Writing – original draft, Writing – review & editing Ahana Gangopadhyay : Conceptualization, Data curation, Software, Writing – review & editing Hsi-Ming Chang : Project administration, Writing – review & editing Ravi Soni : Conceptualization, Funding acquisition, Investigation, Supervision, Writing – review & editing Gopal Avinash : Funding acquisition, Writing – review & editing Ismini Lourentzou : Editor Received 2025 Aug 4; Accepted 2026 Mar 3; Collection date 2026 Apr. © 2026 Ferreira et al This is an open access article distributed under the terms of the Creative Commons Attribution License , which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. PMC Copyright notice PMCID: PMC13061201 PMID: 41950249 Abstract Accurate segmentation of anatomical structures in ultrasound (US) images, particularly small ones, is challenging due to noise and variability in imaging conditions (e.g., probe position, patient anatomy, tissue characteristics and pathology). To address this, we introduce Segment Anything Small (SAS), a simple yet effective scale- and texture-aware data augmentation technique designed to enhance the performance of deep learning models for segmenting small anatomical structures in ultrasound images. SAS employs a dual transformation strategy: (1) simulating diverse organ scales by resizing and embedding organ thumbnails into a black background, and (2) injecting noise into regions of interest to simulate varying tissue textures. These transformations generate realistic and diverse training data without introducing hallucinations or artifacts, improving the model’s robustness to noise and variability. We fine-tuned a promptable foundation model on a controlled organ-specific medical imaging dataset and evaluated its performance on one internal and five external datasets. Experimental results demonstrate significant improvements in segmentation performance, with Dice score gains of up to 0.35 and an average improvement of 0.16 [95% CI: 0.132,0.188]. Additionally, our iterative point prompts provide precise control and adaptive refinement, achieving performance comparable to bounding box prompts with just two points. SAS enhances model robustness and generalizability across diverse anatomical structures and imaging conditions, particularly for small structures, without compromising the accuracy of larger ones. By offering a computationally efficient solution that eliminates the need for extensive human labeling efforts, SAS enables performance improvements in medical image analysis, particularly in resource-constrained settings. Author summary We developed a simple way to help artificial intelligence better outline structures in ultrasound images, especially very small ones that are often difficult to detect accurately. Ultrasound images can vary a lot because of differences in body shape, probe position, image quality, and disease. These changes make it harder for computer models to work reliably across patients and settings. In our study, we created a training approach that makes existing ultrasound images appear at different sizes and with different texture patterns, while still preserving realistic anatomy. This gave the model more varied examples to learn from without creating artificial structures or requiring large amounts of new manual labeling. We found that this approach improved performance across several ultrasound datasets, including challenging cases involving small targets such as follicles, breast tumors, and thyroid nodules. The benefits were strongest when training data were limited, but improvements remained even with much larger datasets. We believe this work shows that carefully designed image transformations can make ultrasound analysis more accurate and robust, which could help support medical imaging tools in settings where expert-labeled data are scarce. Introduction Accurate multi-organ segmentation is critical for applications such as computer-aided diagnosis, surgical planning, navigation, and radiotherapy. However, the task presents significant challenges given the wide variation in organ sizes, ranging from large structures in the chest and abdomen to smaller ones in the head, neck, and gynecological regions. These size disparities often lead to poor segmentation performance, particularly for smaller organs. This, along with high variability of the background region can result in ambiguity in discriminating structure boundaries, followed by the subsequent deterioration of the segmentation performance [ 1 ]. Previous work has experimented with several approaches to improve small organ segmentation performance in the context of expert models that were fine-tuned for segmenting specific organs. Cascaded multi-stage networks have often been used for coarse-to-fine segmentation to iteratively shrink the input region, leading to more accurate segmentation of small anatomical structures [ 2 – 4 ]. Other approaches include incorporating attention mechanisms in U-Net upsampling layers [ 5 ], using additional sub-networks for localization [ 6 ], imposing shape constraints [ 6 ], or employing multi-level structural loss functions that fuse region boundary and pixel-wise information [ 7 ]. However, these methods often require architectural modifications and introduce additional computational overhead. Moreover, many are optimized for specific organs, limiting their generalizability. In contrast, data augmentation provides a model- and anatomy-agnostic alternative to enhancing segmentation performance. Yet, conventional augmentation strategies, such as geometric transformations and intensity perturbations, do not always enhance performance for small organ segmentation, as they may introduce unrealistic variations or fail to sufficiently expand the representation of small structures in training data [ 8 , 9 ]. This highlights the need for specialized augmentation techniques tailored to small organ segmentation in ultrasound imaging. More recently, foundation models trained on large-scale, diverse datasets have demonstrated remarkable success across multiple domains. These models serve as adaptable backbones for various downstream tasks through fine-tuning. The Segment Anything Model (SAM) [ 10 ], for instance, has been fine-tuned on 2D medical images, enabling prompt-based segmentation across different modalities and anatomies [ 11 , 12 ]. However, while these models perform well on general segmentation tasks, small organ segmentation remains an area with significant room for improvement [ 12 , 13 ]. In this work, we address these challenges by introducing Segment Anything Small (SAS), a dual transformation strategy designed to enhance data variability in terms of organ scale and tissue texture. Unlike conventional augmentation techniques, SAS explicitly improves the representation of small structures, making it particularly well-suited for multi-organ ultrasound segmentation. To evaluate its effectiveness, we fine-tuned a lightweight transformer-based promptable foundation model using SAS in two scenarios: a limited single-organ dataset with low diversity and a larger, more diverse multi-organ dataset. Our experimental results, conducted on one internal and five external datasets from diverse organs, demonstrate that SAS significantly improves segmentation performance across both large and small anatomical structures using click prompts. Notably, our results show that fine-tuning medical imaging foundation models with SAS leads to more robust segmentation across varying organ sizes, even when the fine-tuning dataset is relatively small and lacks diversity. Furthermore, SAS remains effective even as the training dataset expands 80-fold (from 1,050 to 80,000 images), demonstrating its robustness in both data-scarce and data-abundant scenarios. In summary, this work makes the following contributions: We propose SAS, a simple yet effective dual transformation strategy that combines image scaling into thumbnails with region-of-interest (ROI) perturbation, enhancing the network’s ability to learn shape and texture information of small organs. SAS is a non-generative data variety approach that generates realistic and semantically consistent perturbed images, circumventing regulatory and compliance concerns associated with generative synthetic data approaches, such as those related to the Health Insurance Portability and Accountability Act (HIPAA) and the General Data Protection Regulation (GDPR), as highlighted in [ 14 ]. This makes SAS a practical and ethically sound solution for medical imaging. Unlike generative methods, SAS prevents the introduction of artifacts, unwanted structures, or hallucinations, making it highly suitable for aiding in the development of AI-based medical applications, where accuracy and reliability are critical. SAS enhances small-organ segmentation performance while maintaining accuracy for larger organs. Through extensive experiments on both small and large organ segmentation tasks in ultrasound images, we compared models trained with and without SAS using the well-benchmarked MedSAM [ 11 ] foundation model. SAS enhances model robustness and generalizability, particularly across out-of-distribution target structures and anatomical regions. By promoting learned invariances and balancing the training set, SAS generates more instances of small structures while preventing overfitting to larger organs. Extensive experiments in both data-scarce and data-abundant scenarios confirm the effectiveness of SAS in improving performance across a wide variety of anatomical structures and datasets. Indeed, the recent success of large foundation models trained on enormous quantities of data prompts a critical question: Can we enhance medical imaging datasets by improving the sufficiency and data diversity of training examples through non-generative transformations? By generating large-scale datasets from a limited number of real images, this approach diminishes the need for human labeling efforts, while ensuring the process remains free from hallucinations. Methods Pre-training a foundation model involves two key steps: collecting a large and diverse dataset, and training a base model that can later be adapted via transfer learning for downstream applications. In this work, we focus on the “last mile” of this process — fine-tuning a foundation model for the task of multi-organ segmentation in ultrasound imaging. To achieve this, we utilize a lightweight version of MedSAM, a model pre-trained on a large-scale medical imaging dataset comprising over 1.5 million images across 10 imaging modalities, including 94,450 ultrasound images. By leveraging this pre-trained foundation model, we harness the power of large-scale medical data without the computational cost and resource burden of training from scratch. Fine-tuning such a foundation model for downstream tasks often yields superior performance, provided sufficient training data is available. Ethics statement This study was reviewed and approved by the GE HealthCare Institutional Review Board. The study involved retrospective analysis of medical imaging data. For internal datasets (Kidney, Gynecology, and Follicle), all data were fully de-identified prior to access, and the requirement for informed consent was waived by the Institutional Review Board due to the retrospective nature of the study and the use of de-identified data. For publicly available datasets used in this work, ethical approval and informed consent procedures were handled by the original dataset providers, in accordance with their respective institutional and regulatory requirements. Segment Anything Small (SAS) image transformations The proposed method employs a dual transformation strategy to simulate various organ scales and appearances, as illustrated in Fig 1 . Step 1 enhances the model’s robustness to scale differences, which is particularly important for accurately segmenting small organs that may appear at different resolutions in real-world ultrasound images. Step 2, through noise injection, further enhances the model’s ability to generalize across diverse tissue appearances while preserving the structural integrity of the original data. Together, these steps improve the model’s performance and adaptability across a wide range of anatomical structures and imaging conditions. Fig 1. SAS image transformations. Open in a new tab Step 1: Thumbnail generation by resizing images to simulate small anatomical structures. Step 2: Region-of-interest perturbations to represent variations in pixel intensity and noise levels. Specifically, given a network input image and its corresponding mask pair ( x , y ), where x represents the image and y denotes the mask, SAS generates a transformed image pair ( x ′, y ′) through a two-step process. First, the ultrasound window region within the image is resized down to generate scaled thumbnails, which are then placed on a black background of size equal to the desired network input size. Second, noise is injected into the organ region defined by the segmentation mask y to simulate diverse textures. Next, we provide a detailed explanation of these steps. Step 1: Simulating organ scales. In the first step, we simulate variations in organ size by extracting the region of interest (ROI) within the ultrasound window and rescaling it to create a thumbnail, as illustrated in Fig 1 (Step 1). This thumbnail is then placed on a black background with the same dimensions as the network’s input image x , resulting in a scaled, zoomed-out representation of the organ. The black background is generated by creating a blank image matching the input dimensions with all pixel values set to zero, ensuring that no additional noise or artifacts are introduced. This operation is designed to mimic the on-screen display behavior of ultrasound systems during zoom in/out, where the ultrasound content occupies a subset of the screen, and the remaining pixels correspond to non-displayed regions outside the rendered ultrasound frame (e.g., outside the sector/convex/trapezoid viewport), which are typically encoded as zero-valued pixels. Using a zero-intensity background therefore preserves the display geometry without introducing artificial texture or speckle patterns. Because these padded regions contain no anatomical signal, they contribute minimal gradients during backpropagation and do not introduce spurious learning signals, keeping parameter updates driven by the displayed ultrasound content. During fine-tuning, the input image size is fixed—for example, at 256 × 256 pixels in our selected architecture, as described in Section Data pre-processing . The thumbnail size is randomly selected during training, ranging from 64 × 64 to 256 × 256 pixels in this implementation, corresponding to a scaling ratio of 0.25 to 1. This simulates organs of varying sizes relative to the input image. Note that the input image size and thumbnail size can be adjusted depending on the neural network architecture. Step 2: Simulating diverse textures via noise injection. In the second step, we simulate diverse textures in organs and tumors by injecting noise into the regions of interest (ROI) defined by the segmentation masks, as illustrated in Fig 1 (Step 2). The perturbation is achieved by applying one of several noise types—Speckle, Gaussian, Salt and Pepper, or Poisson—chosen at random with equal probability. Only one noise type is applied at a time to ensure controlled and realistic variations in texture. During training, 50% of the images containing large structures, defined in the next section, are randomly transformed by SAS. Organ size definition. We classify anatomy size based on its relative size to the original image: segmentation masks covering up to 3% of the image’s original size are defined as “small structures,” or “large structures” otherwise. This classification ensures that small organs or pathological structures (e.g., tumors) are not classified as small if they are zoomed in, i.e., if their actual size within the image exceeds the 3% threshold. To distinguish small from large structures, we classify anatomy based on its relative size with respect to the full image. Specifically, structures whose segmentation masks occupy at most 3% of the image area are defined as small, while larger structures are categorized as large. This threshold was selected empirically based on the distribution of relative structure sizes in our datasets, which exhibits a strong concentration of samples below 3% area and a pronounced tail beyond this value ( Fig 2 ). The 3% cutoff corresponds to a natural transition point in the size distribution, separating the dense cluster of small structures from larger, more variable anatomies. Fig 2. Distribution of relative structure sizes used to define small and large anatomies. Open in a new tab Histogram of the relative area ratio (structure area divided by image area) across all datasets. Model training and selection We adapt the network architecture described in [ 16 , 15 ], which includes a TinyViT image encoder [ 17 ], and a SAM-based prompt encoder and mask decoder [ 11 , 10 ]. The image encoder maps the input image into a high-dimensional embedding space, while the prompt encoder transforms user-click prompts (described in Section Multi-click prompts ) into feature representations using positional encoding [ 18 ]. The mask decoder then fuses the image embeddings and prompt features through cross-attention mechanisms [ 19 ]. The SAS model is initialized using weights from a lightweight version of MedSAM [ 11 ], specifically LiteMedSAM [ 15 ]. LiteMedSAM was created by distilling a smaller image encoder, TinyViT, from the original ViT-based MedSAM image encoder, ensuring alignment in their image embedding outputs. We then fine-tune the model using iterative click prompts, adapting MedSAM’s prompt encoder to support a technique we call the “Multi-click prompting strategy,” presented in Section Multi-click prompts . This strategy simulates realistic user-generated point prompts and is detailed in the following section. During fine-tuning, all trainable parameters in the image encoder, prompt encoder, and mask decoder are updated. The image encoder, prompt encoder, and mask decoder contain 5,726,740, 6,220, and 4,058,340 trainable parameters, respectively, resulting in a total model size of 9,791,300 parameters. For optimization, we use a compound loss function combining Cross-Entropy loss for pixel-wise classification and Dice loss to improve segmentation accuracy [ 20 ]. The network is optimized using the AdamW optimizer [ 21 ] with hyperparameters β 1 = 0.9 , β 2 = 0.999 , an initial learning rate of 5 e − 5 , and a weight decay of 0.01. Training is performed with a batch size of 8, on a single A100 (80 GB) GPU. Early stopping is applied if the combined loss on the validation set does not improve for 5 consecutive epochs. The final model checkpoint is selected based on the highest validation metric value. Multi-click prompts. Point click prompts were used instead of bounding boxes for ultrasound segmentation, as bounding boxes are less effective for non-convex shapes and may inadvertently include unwanted anatomical structures, leading to incorrect segmentation. While previous research [ 22 ] suggests that box-based prompts can yield relatively higher segmentation accuracy, large bounding boxes may encompass multiple instances and background structures, potentially confusing the model and resulting in incorrect segmentation. Other work [ 12 ] supports the use of iterative click prompts for achieving similar results to bounding boxes. We implemented an iterative click prompt technique to simulate realistic user-generated interactions. Initially, the first point is selected from an area near the centroid of the reference standard (RS) mask. Specifically, the ( x , y ) coordinates are chosen from within the 30th percentile of the largest Euclidean distances to the mask boundary. Subsequent points are determined by calculating an error map between the model’s prediction and the RS mask. Points are then selected from regions with the highest error. Each point is labeled as either a positive or negative prompt based on its location relative to the RS mask: a point is classified as a false negative if it falls inside the RS mask (indicating an error region), and as a false positive if it falls outside the RS mask. Data The effectiveness of foundation models is strongly influenced by the size of the dataset, as highlighted by scaling law studies [ 11 , 23 , 24 ], particularly in medical imaging applications, where dataset availability and diversity can significantly impact the generalization capability of the models. For instance, MedSAM [ 11 ] - trained on over 1.5 million medical image-mask pairs across 10 imaging modalities, including 94,450 ultrasound (US) images - investigates the impact of varying the size of the training dataset, and confirms that increasing the number of training images significantly improves results in both internal and external validation sets. However, its evaluation on ultrasound has been limited to a single external dataset, focusing exclusively on fetal head segmentation—an easy-to-segment target where even the original SAM (without medical fine-tuning) achieved a Dice Similarity Coefficient (DSC) of 92.4% [ 11 ]. This underscores the need for a more comprehensive evaluation of segmentation performance on small structures across varied datasets. In this work, we aim to investigate the model’s ability to generalize effectively—leveraging the vast amounts of data seen during training to perform well on tasks involving unseen data, a core principle of foundation models. Specifically, we evaluate the performance of such models on out-of-distribution anatomies—specifically, i.e., datasets containing structures and imaging conditions not encountered during training. Additionally, we assess the feasibility and impact of the proposed SAS approach in improving segmentation performance for small anatomical structures, which are often challenging due to their size and variability. To achieve this, we propose two evaluation scenarios: Scenario 1 investigates the model’s ability to generalize to new, out-of-distribution anatomies using a limited training dataset consisting of 1,050 grayscale images from an internal dataset. These images depict both longitudinal and transverse views of renal anatomy, with each image being a single frame. The training data was randomly split per patient into 80% for training and 20% for validation, with the objective of segmenting the kidney. Scenario 2 examines the impact of increasing the training dataset size by incorporating a larger variety of ultrasound image data, consisting of 87,000 images of gynecological structures. As in Scenario 1, the data in Scenario 2 was split into 80% for training and 20% for validation, with the goal of segmenting these gynecological anatomical structures. In both scenarios, we evaluate the same test sets, including small anatomies such as breast tumors, thyroid, nerve, vessels, gallbladder, and ovarian follicles. In Scenario 1, all these anatomies are considered out-of-distribution target structures, while in Scenario 2, only the anatomy of the ovarian follicle becomes part of the in-distribution target structure data. The test set includes both an internal follicle dataset and external publicly available medical image segmentation datasets, obtained from the Cancer Imaging Archive [ 25 ] and Kaggle, as detailed in Table 1 . These public datasets are well-established in the literature and include reference standard annotations provided by human experts. None of the holdout test datasets were ever used during model fine-tuning for either training or validation. In the following, we present each dataset and its role in model development. Table 1 summarizes patient demographics and key dataset characteristics. Table 1. Ultrasound datasets used for training and testing. Dataset Used for Nb. masks Unique patients Object of interest Demographics/clinical info Kidney train 1,050 1,050 Kidney Not available Gynecology train 80,000 920 Uterus Not available Follicle test 69,000 71 Follicle Not available Abdominal [ 26 ] test 42 11 Kidney, vessels, gallbladder 40% female, age = 27 ± 3 years; no history of abdominal pathology or known disease. BUSI [ 27 ] test 665 600 Breast benign and malignant tumors 100% female, age 25–75 years; 454 benign and 211 malignant cases. BrEaST [ 25 ] test 264 256 Breast tumor 100% female, age 18–87 years; normal, benign, and malignant categories. Histological diagnoses are reported in [ 25 ]. DDTI [ 28 ] test 656 299 Thyroid 90% female, age = 57 ± 16 years; lesions include adenomas, thyroiditis, cystic nodules, and thyroid cancers. Neck [ 29 ] test 2323 N/A Nerve structures Not available. Open in a new tab Nb. denotes number. N/A indicates not available. Data pre-processing. The ultrasound region of interest was extracted from DICOM images, using the DICOM metadata and converted to PNG format. Next, the images were resized at the longest dimension to align with the input size of the model’s encoder, i.e., 256 pixels, using bilinear interpolation. We then apply min-max normalization from 0 to 1, and pad the resized images with zero values, as needed, to create square dimensions of 256 × 256 × 3. The preprocessing for reference standard masks is similar, with a few key differences. Instead of bilinear interpolation, we use nearest-exact interpolation to preserve the mask’s pixel values. After resizing the masks to match the input size of the image encoder, we pad them with zeros to achieve square dimensions of 256 × 256 × 3. Quantitative metrics The Dice Similarity Coefficient (DSC) [ 30 ] and Normalized Surface Distance (NSD) [ 30 ], are widely used for evaluation in medical image segmentation tasks. Both metrics, range from the smallest value of 0 indicating the worst performance, to the highest value of 1.0 indicating perfect alignment. DSC is a measure of overlap between the predicted segmentation P and the reference standard segmentation R . It is defined as: DSC = 2 | 𝐏 ∩ 𝐑 | | 𝐏 | + | 𝐑 | where | 𝐏 ∩ 𝐑 | represents the intersection of the predicted and reference standard segmentation, and | P | and | R | are the cardinalities of the predicted and reference standard sets, respectively. The NSD is an uncertainty-aware segmentation metric that measures the overlap between two boundaries. It estimates which fraction of a segmentation boundary is correctly predicted with an additional threshold τ related to the clinically acceptable amount of deviation in pixels [ 31 ], and is defined as: NSD = | S P ∩ β R ( τ ) | + | S R ∩ β P ( τ ) | | S P | + | S R | where S P is the boundary of the predicted segmentation P , β G ( τ ) is the border regions of R at tolerance τ , S R is the boundary of the reference standard segmentation R , β P ( τ ) is the border regions of P at tolerance τ , and τ is the degree of strictness for what constitutes a correct boundary. Unlike DSC, which focuses on area overlap, NSD evaluates the accuracy of the segmentation boundaries, making it particularly sensitive to errors in the shape and contour of small objects [ 30 ]. It is thus a measure of what fraction of a segmentation boundary would have to be redrawn to correct for segmentation errors. Results We conducted experiments in two scenarios, as detailed in Section Data . Scenario 1 uses a limited dataset of 1,050 renal ultrasound images to train the model, while Scenario 2 incorporates a larger dataset of 87,000 gynecological ultrasound images to study the impact of dataset size and diversity. Both scenarios are evaluated on the same test set, which includes small anatomical structures such as breast tumors, thyroid, nerves, vessels, gallbladder, and ovarian follicles. Iterative click prompt We begin by evaluating the effectiveness of iterative click prompts for segmentation performance. Iterative click prompts were generated as described in Section Multi-click prompts , and this strategy was applied during both training and inference. Performance was assessed in Scenario 1 on a holdout test split of the kidney dataset (n = 126) using Dice scores across the number of iterative prompts used, ranging from 1 to 10. Evaluations were conducted under two different fine-tuning scenarios for iterative click prompting: (a) full fine-tuning, where the image encoder, mask decoder, and prompt encoder were all fine-tuned, and (b) partial fine-tuning, where the image encoder was frozen while only the mask decoder and prompt encoder were fine-tuned. Additionally, a bounding box prompt approach with full fine-tuning was evaluated on the same dataset for comparison. As shown in Fig 3 , the iterative click prompts achieve comparable performance to bounding box prompts with just 2 click prompts, aligning with previous findings that iterative click prompts can surpass bounding boxes as the number of clicks increases [ 12 ]. The results also reveal a significant difference between full and partial fine-tuning. Full fine-tuning consistently achieved higher Dice scores with fewer prompts when compared to partial fine-tuning, and therefore, we adopt this setup for all subsequent experiments in this section. Notably, the full fine-tuning approach shows performance comparable to bounding box prompts in as few as 2 iterative click prompts, highlighting its efficiency and effectiveness in achieving high-quality segmentation. Fig 3. Segmentation performance: iterative point vs. bounding box prompts. Open in a new tab Comparison of segmentation performance using iterative point and bounding box prompts with full and partial fine-tuning. Dice scores are reported for iterative point prompts ranging from 1 to 10. Visualizing data diversity To better understand the impact of SAS on data diversity, we computed feature embeddings using the ImageNet-1K [ 32 ] pre-trained VGG-16 model for both original ultrasound images and SAS-generated images. These embeddings were visualized using t-SNE plots in both Scenario 1 and Scenario 2, as shown in Fig 4 . The plots reveal well-separated clusters for different test sets, underscoring the diversity of the evaluation datasets. In Scenario 1, where the original dataset is severely limited (1,050 renal ultrasound images), the original data points form a tight cluster (light orange), indicating restricted variability. SAS-generated data points (dark orange), created by applying SAS 20 times per image, occupy distinct regions of the data subspace, expanding the feature space, and improving the model’s ability to generalize. In Scenario 2, due to computational constraints and for the sake of improving visualization, we randomly down-sampled the gynecological training dataset to 20,000 images and applied SAS once per image. Here, SAS-generated points also enhance data variety, though the effect is less pronounced compared to Scenario 1 due to the high diversity of the original data itself. Fig 4. Two-dimensional t-SNE visualization of feature embeddings for SAS-generated images, training set images, and test set images in Scenario 1 (left) and Scenario 2 (right). Open in a new tab SAS-generated images in both scenarios total 20,000, while non-SAS images in Scenario 2 were downsampled to 20,000 for comparison. Test set details are provided in Table 1 . The embeddings were extracted using a VGG-16 encoder pre-trained on ImageNet-1K, with a t-SNE perplexity of 50. While the t-SNE plots provide meaningful insights into data diversity, it is important to note that these plots help us visualize data points reduced to two dimensions, so the plots may not fully capture the complexity of the data. A more complete picture regarding the effectiveness of SAS in enhancing data diversity can be obtained by also considering the improved generalization performance demonstrated in the next section. Quantitative performance evaluation of small structures segmentation Building on the insights from the t-SNE analysis, which demonstrated SAS’s ability to enhance data diversity and improve generalization, we now quantitatively evaluate its impact on segmenting anatomical structures in ultrasound. Figure 5 presents the Normalized Surface Distance (NSD) and Dice Similarity Coefficient (DSC) results across all six test datasets (detailed in Table 1 ) under two training scenarios: (a) Scenario 1 (orange lines), trained on a small, low-variety kidney image dataset, and (b) Scenario 2 (blue lines), trained on a large, high-variety gynecological image dataset. Performance is compared between models trained with and without SAS, highlighting its impact on segmentation accuracy across diverse anatomical structures. SAS consistently improves performance across both small and large organs, with particularly notable gains for small structures, where scale variations pose a greater challenge. Fig 5. Evaluation of SAS. Open in a new tab Normalized Surface Distance (NSD) and Dice Similarity Coefficient (DSC) results for all test datasets (see Table 1 ) in two scenarios: a) Scenario 1 (orange lines), using a low-variety training dataset of kidney images, and b) Scenario 2 (blue lines), using a high-variety training dataset of gynecological images. Results are shown for models trained with and without SAS. Numbers 1 to 8 indicate the number of iterative click prompts used in the segmentation process. Confidence intervals (CIs) were bootstrapped 10,000 times. Asterisks (*) indicate that baseline non-SAS and SAS are not significantly different. Notably, SAS in Scenario 1 (orange solid line) achieves performance levels comparable to the more diverse training set in Scenario 2 (blue solid line) for many test sets, demonstrating its ability to enhance generalization even with limited training data. Quantitatively, SAS improves average DSC by 0.084 [95% CI: 0.056, 0.111] in Scenario 1 and by 0.043 [95% CI: 0.022, 0.068] in Scenario 2, averaged across all test sets and eight click-prompt scenarios. When focusing on small organs, SAS achieves an average DSC improvement of 0.16 [95% CI: 0.132, 0.188] in Scenario 1 and 0.035 [95% CI: 0.018, 0.052] in Scenario 2, averaged across all test sets and eight click-prompt scenarios. Across datasets, SAS demonstrates substantial improvements. For instance, in Fig 5(d) , within the BUSI dataset, SAS achieves a DSC improvement of up to 0.35 (Scenario 1, one-click prompt), with an average DSC gain of 0.253 [95% CI: 0.216, 0.289] across eight click-prompt scenarios. The BrEaST dataset shows the largest DSC improvement of 0.161 for a single-click prompt, with an average increase of 0.101 [95% CI: 0.079, 0.125] over eight clicks. In the Abdomen dataset, SAS improves DSC by up to 0.307 for a single-click prompt, with an average gain of 0.234 [95% CI: 0.202, 0.268]. For the Thyroid dataset, the best improvement is observed with a two-click prompt (DSC increase of up to 0.185), with an average improvement of 0.133 [95% CI: 0.104, 0.16]. The Nerve dataset exhibits a maximum DSC improvement of 0.095, with an average DSC gain of 0.04 [95% CI: -0.02, 0.087]. The limited improvement observed on the nerve dataset is consistent with prior findings that nerve segmentation in ultrasound is primarily constrained by the need for long-range anatomical context and preservation of thin, deformable structures, which are not explicitly addressed by scale-based augmentation alone [ 33 ]. In the follicle test set, while SAS significantly boosts average DSC in Scenario 1 by 0.256 [95% 0.234, 0.258) points for small structures, it does not fully bridge the performance gap relative to Scenario 2 (blue lines). This suggests that although SAS enhances segmentation robustness, it cannot entirely compensate for the inherent limitations of the lower-diversity kidney dataset. The violin plots in Figs 6 – 8 show the distribution of NSD and DSC metrics, grouping results from SAS and Non-SAS models while showing individual data points for all six test sets. As before, Scenario 1 is trained on a low-diversity dataset of kidney images, while Scenario 2 leverages a more diverse set of gynecological organs. Ideally, a generalizable segmentation model should achieve high performance with minimal labeled examples from the target anatomy, and these visualizations provide insight into how SAS influences segmentation robustness across different dataset compositions. Fig 6. Inference on Follicle and Breast Tumor - Violin plots showing the dice similarity coefficient and normalized surface distance scores for the small structures follicle and breast tumor datasets follicle (a-b) and BUSI (c-d) respectively. Open in a new tab The training dataset for Scenario 1 consists of kidney images, while Scenario 2 uses gynecological images. Fig 8. Inference in BrEaST and Thyroid test sets. Open in a new tab Violin plots showing the DSC and NSD scores for the BrEaST (a-b) and DDTI (Thyroid) (c-d) datasets. The training dataset for Scenario 1 consists of kidney images, while Scenario 2 uses gynecological images. Fig 7. Inference on Abdomen and Nerve - Violin plots showing the NSD and DSC scores for the Abdomen (a-b) and Neck (c-d) datasets. Open in a new tab The training dataset for Scenario 1 consists of kidney images, while Scenario 2 uses gynecological images. The Non-SAS model shows a higher concentration of low DSC and NSD values compared to the SAS model, indicating poorer segmentation quality, particularly for small anatomical structures. In most settings, the Non-SAS model achieves satisfactory performance only on larger organs (green dots in the plot), whereas SAS demonstrates strong segmentation improvements even with minimal user interaction. The follicle dataset in Fig 6a , 6b serves as a representative case for small organs due to their inherently small size — illustrated by the red points in the violin plots — where 90% of the masks fall within our defined small-organ category, as detailed in Section Organ Size Definition . This enhancement brings performance closer to that observed in Scenario 2, which serves as a best-case benchmark and upper bound for the model’s performance on the follicle dataset. As expected, when training is conducted in Scenario 2, the performance gap between SAS and Non-SAS narrows. However, SAS continues to provide added value, particularly with a limited number of interactive clicks. Qualitative performance evaluation of small structures segmentation Figure 9 presents segmentation examples for three cases from the folicle testset, comparing the reference standard masks with predictions from both the SAS and non-SAS approaches using models trained in Scenario 1. The non-SAS approach struggles with small structures, often failing to segment them entirely, while also exhibiting under-segmentation in larger structures. In contrast, the SAS approach demonstrates improved accuracy across different anatomical sizes, effectively capturing small structures while also providing more complete segmentation for larger ones. Fig 9. Visualized segmentation examples of non-SAS and SAS segmentation masks on the follicle dataset. Open in a new tab The quantitative results presented in Section Quantitative performance evaluation of small structures segmentation and qualitative results in Fig 9 highlight that, despite the limited fine-tuning dataset size and distinctly different structural and anatomical characteristics in the test set, SAS effectively accomplishes the segmentation of small organs and structures, which also underscores the importance of methods for increasing data variety for developing robust models. The ability to generalize across various anatomical structures, despite being trained on a different dataset, demonstrates the adaptability and effectiveness the proposed method. Ablation study: effect of scale and noise perturbations To assess the individual and combined contributions of scale perturbation and noise perturbation in our SAS strategy, we conducted an ablation study in which the model was trained on the Kidney dataset and evaluated on the corresponding test sets. Performance is measured using the Dice coefficient as a function of the number of user interactions (1–8 clicks). To further analyze the impact of SAS under different anatomical conditions, we stratify the results by object size into small and large organs. Figure 10 summarizes the results averaged across the test datasets. Each curve represents the mean Dice score, with shaded regions indicating the standard deviation across datasets. Fig 10. Ablation of scale and noise perturbations in SAS. Open in a new tab Mean Dice score as a function of the number of user clicks (1–8), averaged across all test datasets for small (left) and large organs (right) for No SAS (baseline), Noise only, Scale only, and SAS (Scale + Noise). Shaded regions denote the standard deviation across datasets. For small structures ( Fig 10 , left), the baseline model exhibits limited improvement as the number of clicks increases, saturating at a relatively low Dice score. Introducing noise perturbation alone improves robustness during early interactions, while scale perturbation alone yields larger gains as additional clicks are provided. The full SAS strategy consistently outperforms all ablated variants across all interaction regimes, achieving both faster convergence at early clicks and higher final accuracy. For large structures ( Fig 10 , right), all methods benefit substantially from increased user interactions. Scale perturbation provides clear improvements over the baseline, especially at intermediate and later clicks, whereas noise perturbation alone yields more modest gains. The combined SAS strategy achieves the highest Dice scores across all interaction levels. These results highlight the complementary nature of scale and noise perturbations, particularly for challenging small anatomical targets. Discussion and conclusions In this work, we address the challenge of accurate segmentation in Ultrasound imaging, particularly for small anatomical structures, where data scarcity and limited annotation quality often hinder the generalization of deep learning models. Segment Anything Small (SAS) is easily applicable to any US image, as it does not rely on external segmentations or separately trained generative models. Unlike generative methods, SAS avoids the introduction of artifacts, unwanted structures, or hallucinations, making it highly suitable for medical applications where accuracy and reliability are critical. The proposed approach is based on the assumption that the “expected” domain shift for a specific anatomy can be simulated by applying SAS to a single source domain, enabling a deep model trained with SAS to better generalize to unseen domains. SAS demonstrates significant performance improvements across all six datasets - one internal and five external - for seven segmentation tasks, namely ovarian follicles, breast tumor, thyroid, gallbladder, vessels, kidney, and nerve. Despite being trained on a small, kidney-only dataset lacking geometric feature diversity (Scenario 1), SAS generalizes well to unseen organs. It performs consistently across datasets from multiple centers with diverse imaging vendors and protocols, highlighting its ability to adapt to variations in small organs. In the Large Structures of Scenario 1, however, SAS underperforms for large breast tumors in the BrEaST dataset ( Fig 5a , 5c ), orange lines. The mean Dice score for large organs decreases from approximately 0.63 (no-SAS) to 0.41 (SAS) at one click, corresponding to a reduction of approximately 0.22 Dice. This gap persists across interaction levels, with differences of approximately 0.15 at three clicks and 0.10 at five clicks, before narrowing to about 0.03 at eight clicks. Similar trends are observed in NSD, indicating that degradation is systematic in both overlap- and boundary-based metrics rather than driven by isolated cases. This is likely due to SAS’s inability to capture the morphological and anatomical irregularities of large tumors after altering the training data distribution with artificially generated small target samples. Additionally, the limited kidney dataset, which also involves a large organ, may have caused the non-SAS models to overfit to large organs, as these models do not benefit from the regularization introduced by SAS. Adding more data, as in Scenario 2, seems to address the overfitting issue where SAS consistently proves beneficial regardless of training set size. When dataset variability is increased by expanding the dataset size by a factor of 80 in Scenario 2 (data abundance), the impact of SAS diminishes but remains significant compared to non-SAS, with the breast segmentation task showing the most notable improvement. We attribute this to two key phenomena: SAS promotes learned invariances that enhance generalization, and it balances the training set distribution by generating additional cases of small structures while avoiding overfitting to large organs. Through extensive experiments in both data-scarce and data-abundant scenarios, we demonstrate that SAS improves robustness and generalizability across out-of-distribution organs and diverse datasets. Future work will examine the applicability of SAS to additional imaging modalities and assess its integration into clinical workflows to evaluate feasibility and generalizability in real-world medical scenarios. Because increasing data diversity through augmentation does not necessarily lead to improved performance [ 8 ], future studies should systematically analyze the effect of augmentation strength by varying parameters such as image scaling, noise intensity, and the probability of applying perturbations during training. In addition, we will investigate the inclusion of elastic deformation-based augmentations to better capture the geometric variability of elongated, deformable structures, such as nerves, and assess whether such transformations complement scale-based augmentation. Finally, the behavior of SAS will be evaluated within self-supervised and semi-supervised learning frameworks to study its effectiveness under reduced annotation regimes. Supporting information S1 Fig. Graphical_abstract. (PDF) pdig.0001309.s001.pdf (1.3MB, pdf) Acknowledgments The authors would like to acknowledge GE HealthCare for providing institutional support for this research. No external funding or specific grant was received for this study. A preliminary version of this work was published in the SPIE Proceedings of Medical Imaging [ 34 ]. Data Availability The datasets used in this study are summarized in Table 1 and include both publicly available and internal datasets. Public datasets used for performance evaluation are openly accessible and are referenced in Table 1 . Additional datasets, including Kidney, Gynecology, and Follicle, are internal, non-public datasets and are not publicly available due to protected patient health information and institutional data governance restrictions. These datasets were used for model training (Kidney and Gynecology) and evaluation (Follicle). Access to these data may be considered upon reasonable request to the HealthCare Real World Data (RWD) Team at: [email protected] , subject to institutional approval and applicable data-sharing agreements. Funding Statement The author(s) received no specific funding for this work. References 1. Liu X, Qu L, Xie Z, Zhao J, Shi Y, Song Z. Towards more precise automatic analysis: a systematic review of deep learning-based multi-organ segmentation. Biomed Eng Online. 2024;23(1):52. doi: 10.1186/s12938-024-01238-8 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 2. Zhou Y, Xie L, Shen W, Wang Y, Fishman EK, Yuille AL. A fixed-point model for pancreas segmentation in abdominal CT scans. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. 2017. p. 693–701. 3. Roth HR, Shen C, Oda H, Sugino T, Oda M, Hayashi Y, et al. A Multi-scale pyramid of 3D Fully convolutional networks for abdominal multi-organ segmentation. In: Lecture Notes in Computer Science. Springer International Publishing. 2018. p. 417–25. 10.1007/978-3-030-00937-3_48 [ DOI ] [ Google Scholar ] 4. Yu Q, Xie L, Wang Y, Zhou Y, Fishman EK, Yuille AL. Recurrent saliency transformation network: incorporating multi-stage visual cues for small organ segmentation. In: 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2018. 8280–9. 10.1109/cvpr.2018.00864 [ DOI ] 5. Schlemper J, Oktay O, Schaap M, Heinrich M, Kainz B, Glocker B, et al. Attention gated networks: learning to leverage salient regions in medical images. Med Image Anal. 2019;53:197–207. doi: 10.1016/j.media.2019.01.012 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 6. Gao Y, Huang R, Yang Y, Zhang J, Shao K, Tao C, et al. FocusNetv2: imbalanced large and small organ segmentation with adversarial shape constraint for head and neck CT images. Med Image Anal. 2021;67:101831. doi: 10.1016/j.media.2020.101831 [ DOI ] [ PubMed ] [ Google Scholar ] 7. Liu Y, Duan Y, Zeng T. Learning multi-level structural information for small organ segmentation. Signal Process. 2022;193:108418. doi: 10.1016/j.sigpro.2021.108418 [ DOI ] [ Google Scholar ] 8. Tupper A, Gagné C. Analyzing data augmentation for medical images: a case study in ultrasound images. arXiv. 2024. https://doi.org/arXiv:2403.09828 9. Sfakianakis C, Simantiris G, Tziritas G. GUDU: geometrically-constrained ultrasound data augmentation in U-Net for echocardiography semantic segmentation. Biomed Signal Process Control. 2023;82:104557. doi: 10.1016/j.bspc.2022.104557 [ DOI ] [ Google Scholar ] 10. Kirillov A, Mintun E, Ravi N, Mao H, Rolland C, Gustafson L, et al. Segment Anything. In: 2023 IEEE/CVF International Conference on Computer Vision (ICCV). 2023. 3992–4003. 10.1109/iccv51070.2023.00371 [ DOI ] 11. Ma J, He Y, Li F, Han L, You C, Wang B. Segment anything in medical images. Nat Commun. 2024;15(1):654. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 12. Cheng J, Ye J, Deng Z, Chen J, Li T, Wang H. SAM-Med2D. arXiv. 2023. https://doi.org/arXiv:2308.16184 13. Mazurowski MA, Dong H, Gu H, Yang J, Konz N, Zhang Y. Segment anything model for medical image analysis: an experimental study. Med Image Anal. 2023;89:102918. doi: 10.1016/j.media.2023.102918 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 14. Templin T, Perez MW, Sylvia S, Leek J, Sinnott-Armstrong N. Addressing 6 challenges in generative AI for digital health: a scoping review. PLOS Digit Health. 2024;3(5):e0000503. doi: 10.1371/journal.pdig.0000503 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 15. Liu W, Bian R, Li H, Tian T, et al. LiteMedSAM with low-rank adaptation and multi-box efficient inference for medical image segmentation. In: CVPR 2024: Segment Anything in Medical Images on Laptop. 2024. 16. Zhang C, Han D, Qiao Y, Kim JU, Bae SH, Lee S. Faster segment anything: towards lightweight SAM for mobile applications. arXiv. 2023. https://doi.org/arXiv:2306.14289 17. Wu K, Zhang J, Peng H, Liu M, Xiao B, Fu J, et al. TinyViT: fast pretraining distillation for small vision transformers. In: European Conference on Computer Vision. 2022. p. 68–85. 18. Tancik M, Srinivasan P, Mildenhall B, Fridovich-Keil S, Raghavan N, Singhal U. Fourier features let networks learn high frequency functions in low dimensional domains. Adv Neural Inf Process Syst. 2020;33:7537–47. [ Google Scholar ] 19. Vaswani A. Attention is all you need. Adv Neural Inf Process Syst. 2017. 20. Isensee F, Jaeger PF, Kohl SAA, Petersen J, Maier-Hein KH. nnU-Net: a self-configuring method for deep learning-based biomedical image segmentation. Nat Methods. 2021;18(2):203–11. doi: 10.1038/s41592-020-01008-z [ DOI ] [ PubMed ] [ Google Scholar ] 21. Loshchilov I, Hutter F. Decoupled weight decay regularization. In: International Conference on Learning Representations. 2019. 22. Wu J, Ji W, Liu Y, Fu H, Xu M, Xu Y. Medical SAM adapter: adapting segment anything model for medical image segmentation. arXiv. 2023. https://doi.org/arXiv:2304.12620 [ DOI ] [ PubMed ] 23. Kaplan J, McCandlish S, Henighan T, Brown TB, Chess B, Child R. Scaling laws for neural language models. arXiv. 2020. https://doi.org/arXiv:2001.08361 24. Vorontsov E, Bozkurt A, Casson A, Shaikovski G, Zelechowski M, Severson K, et al. A foundation model for clinical-grade computational pathology and rare cancers detection. Nat Med. 2024;30(10):2924–35. doi: 10.1038/s41591-024-03141-0 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 25. Pawłowska A, Ćwierz-Pieńkowska A, Domalik A, Jaguś D, Kasprzak P, Matkowski R, et al. Curated benchmark dataset for ultrasound based breast lesion analysis. Sci Data. 2024;11(1):148. doi: 10.1038/s41597-024-02984-z [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 26. Vitale S, Orlando JI, Iarussi E, Larrabide I. Improving realism in patient-specific abdominal ultrasound simulation using CycleGANs. Int J Comput Assist Radiol Surg. 2020;15(2):183–92. doi: 10.1007/s11548-019-02046-5 [ DOI ] [ PubMed ] [ Google Scholar ] 27. Al-Dhabyani W, Gomaa M, Khaled H, Fahmy A. Dataset of breast ultrasound images. Data Brief. 2019;28:104863. doi: 10.1016/j.dib.2019.104863 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 28. Pedraza L, Vargas C, Narváez F, Durán O, Muñoz E, Romero E. An open access thyroid ultrasound image database. In: 10th International Symposium on Medical Information Processing and Analysis. vol. 9287; 2015. p. 188–193. [ Google Scholar ] 29. Baby M, Jereesh AS. Automatic nerve segmentation of ultrasound images. In: 2017 International conference of Electronics, Communication and Aerospace Technology (ICECA). 2017. 107–12. 10.1109/iceca.2017.8203654 [ DOI ] 30. Maier-Hein L, Reinke A, Godau P, Tizabi MD, Buettner F, Christodoulou E, et al. Metrics reloaded: recommendations for image analysis validation. Nat Methods. 2024;21(2):195–212. doi: 10.1038/s41592-023-02151-z [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 31. Seidlitz S, Sellner J, Odenthal J, Özdemir B, Studier-Fischer A, Knödler S, et al. Robust deep learning-based semantic organ segmentation in hyperspectral images. Med Image Anal. 2022;80:102488. doi: 10.1016/j.media.2022.102488 [ DOI ] [ PubMed ] [ Google Scholar ] 32. Deng J, Dong W, Socher R, Li L-J, Kai Li, Li Fei-Fei. ImageNet: a large-scale hierarchical image database. In: 2009 IEEE Conference on Computer Vision and Pattern Recognition. 2009. pp. 248–55. 10.1109/cvpr.2009.5206848 [ DOI ] 33. Liu F, Liu C, Wang L, Ma L, Lu ZM. Segmentation of nerve on ultrasound images using deep adversarial network. Med Phys. 2018. [ Google Scholar ] 34. Ferreira D, Gangopadhyay A, Chang H-M, Soni R, Avinash G. Segment anything small for ultrasound: enhanced segmentation of small anatomical structures using iterative point prompts and image transformations. In: Medical Imaging 2025: Image Processing. 2025. pp. 52. 10.1117/12.3046538 [ DOI ] PLOS Digit Health. doi: 10.1371/journal.pdig.0001309.r001 Decision Letter 0 Ismini Lourentzou Ismini Lourentzou Section Editor Find articles by Ismini Lourentzou Author information Copyright and License information Roles Ismini Lourentzou : Section Editor © 2026 Ismini Lourentzou This is an open access article distributed under the terms of the Creative Commons Attribution License , which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. PMC Copyright notice 23 Dec 2025 Response to Reviewers '. This file does not need to include responses to any formatting updates and technical items listed in the 'Journal Requirements' section below.'. This file does not need to include responses to any formatting updates and technical items listed in the 'Journal Requirements' section below.-->-->* A marked-up copy of your manuscript that highlights changes made to the original version. You should upload this as a separate file labeled ' Revised Manuscript with Track Changes '.'.-->-->* An unmarked version of your revised paper without tracked changes. You should upload this as a separate file labeled ' Manuscript '.'.-->--> -->-->If you would like to make changes to your financial disclosure, competing interests statement, or data availability statement, please make these updates within the submission form at the time of resubmission. Guidelines for resubmitting your figure files are available below the reviewer comments at the end of this letter.-->--> -->-->We look forward to receiving your revised manuscript.-->--> -->-->Kind regards,-->--> -->-->Ismini Lourentzou-->-->Section Editor-->-->PLOS Digital Health-->--> -->-->Ismini Lourentzou-->-->Section Editor-->-->PLOS Digital Health-->--> -->-->Leo Anthony Celi-->-->Editor-in-Chief-->-->PLOS Digital Health-->-->orcid.org/0000-0001-6712-6626-->--> -->--> -->--> Journal Requirements: -->--> -->--> 1. Please send a completed 'Competing Interests' statement, including any COIs declared by your co-authors. If you have no competing interests to declare, please state "The authors have declared that no competing interests exist". Otherwise please declare all competing interests beginning with the statement "I have read the journal's policy and the authors of this manuscript have the following competing interests:" 2. Please ensure that your Ethics Statement is available in its entirety at the beginning of your Methods section, under a subheading 'Ethics Statement'. It must include: i) The name(s) of the Institutional Review Board(s) or Ethics Committee(s) ii) The approval number(s), or a statement that approval was granted by the named board(s) iii) (for human participants or donors) - A statement that formal consent was obtained (must state whether verbal/written) OR the reason consent was not obtained (e.g., anonymity). 3. We ask that a manuscript source file is provided at Revision. Please upload your manuscript file as a .doc, .docx, .rtf or .tex. 4. Please provide separate figure files in .tif or .eps format. For more information about figure files please see our guidelines: https://journals.plos.org/digitalhealth/s/figures https://journals.plos.org/digitalhealth/s/figures#loc-file-requirements 5. We note that you have included your Figures within the body of your manuscript. Please remove the Figures from the body of your manuscript and upload them as separate Figure files. 6. In the online submission form, you indicated that “The datasets used for performance evaluation are publicly accessible and are referenced in Table 1. Additional datasets used in this study, including Kidney, Gynecology, and Follicle, are not publicly available due to protected patient health information. For further inquiries, please contact the Corresponding Author.”. All PLOS journals now require all data underlying the findings described in their manuscript to be freely available to other researchers, either 1. In a public repository, 2. Within the manuscript itself, or 3. Uploaded as supplementary information. This policy applies to all data except where public deposition would breach compliance with the protocol approved by your research ethics board. If your data cannot be made publicly available for ethical or legal reasons (e.g., public availability would compromise patient privacy), please explain your reasons by return email and your exemption request will be escalated to the editor for approval. Your exemption request will be handled independently and will not hold up the peer review process, but will need to be resolved should your manuscript be accepted for publication. One of the Editorial team will then be in touch if there are any issues. Additional Editor Comments (if provided): -->--> -->--> -->--> -->-->[Note: HTML markup is below. Please do not edit.]-->--> -->--> Reviewers' Comments: -->--> -->-->Reviewer's Responses to Questions Comments to the Author 1. Does this manuscript meet PLOS Digital Health’s publication criteria ? Is the manuscript technically sound, and do the data support the conclusions? The manuscript must describe methodologically and ethically rigorous research with conclusions that are appropriately drawn based on the data presented.? Is the manuscript technically sound, and do the data support the conclusions? The manuscript must describe methodologically and ethically rigorous research with conclusions that are appropriately drawn based on the data presented.-->?> Reviewer #1: Yes Reviewer #2: Yes ********** 2. Has the statistical analysis been performed appropriately and rigorously?-->?> Reviewer #1: Yes Reviewer #2: Yes ********** 3. Have the authors made all data underlying the findings in their manuscript fully available (please refer to the Data Availability Statement at the start of the manuscript PDF file)??> The PLOS Data policy requires authors to make all data underlying the findings described in their manuscript fully available without restriction, with rare exception. The data should be provided as part of the manuscript or its supporting information, or deposited to a public repository. For example, in addition to summary statistics, the data points behind means, medians and variance measures should be available. If there are restrictions on publicly sharing data—e.g. participant privacy or use of data from a third party—those must be specified.requires authors to make all data underlying the findings described in their manuscript fully available without restriction, with rare exception. The data should be provided as part of the manuscript or its supporting information, or deposited to a public repository. For example, in addition to summary statistics, the data points behind means, medians and variance measures should be available. If there are restrictions on publicly sharing data—e.g. participant privacy or use of data from a third party—those must be specified.--> Reviewer #1: Yes Reviewer #2: Yes ********** 4. Is the manuscript presented in an intelligible fashion and written in standard English??> Reviewer #1: Yes Reviewer #2: Yes ********** Reviewer #1: This article investigates and proposes a method called SAS for data augmentation, with the research demonstrating that the SAS-based approach effectively enhances the performance of image segmentation, particularly showing greater effectiveness for smaller targets. The overall structure of the article is clear and well-supported by evidence. However, there are a few specific areas where additional clarification or improvement could aid reader comprehension: 1. In Figure 3, Scenarios 1 and 2 use colors that are too similar to represent SAS and no-SAS samples, making comparison challenging. Since Section 3.2 specifically discusses how no-SAS samples cluster tightly in Scenario 1, using more distinct colors would better highlight the separation between SAS and no-SAS distributions as observed through t-SNE visualization. 2. The SAS method achieves data augmentation through two steps: image scaling and noise addition. While both steps contribute positively—scaling improves small target detection and noise enhances robustness—the combined effect shown in Figure 3 via t-SNE doesn't clarify whether the dimensionality reduction reflects the impact of scaling, noise addition, or both. A visualization comparing feature distributions after each step separately would better illustrate their individual contributions. 3. The four types of noise (Speckle, Gaussian, Salt and Pepper, Poisson) used in Step 2 do correspond to typical ultrasound noises. However, in Figure 1, the image added at the bottom of Step 2 appears to be a mosaic image rather than one of the specified noise types. This inconsistency should be reviewed by the authors. The article makes significant contributions but these specific refinements would strengthen its communication of findings. Reviewer #2: 1. phrases such as “emerges as a powerful tool” read as promotional; please revise to reflect performance association rather than definitive superiority. 2. Justify the 3% threshold for “small structures” (Methods, Section 2.1.3, lines 149–154): Please provide a citation or empirical rationale for this cutoff, or include a sensitivity analysis showing robustness to alternative thresholds. 3. Expand justification for black background placement (Methods, Section 2.1.1, lines 131–134): Please explain why a zero-intensity background best reflects ultrasound acquisition physics and whether this risks domain shift. 4. The manuscript would benefit from explicitly stating whether scale and noise components were tested independently, and where those results are reported 5. The explanation for underperformance on large breast tumors is reasonable, but please add a quantitative comparison to show the magnitude and consistency of degradation 6. The nerve dataset shows minimal improvement; please add a brief explanation of why SAS may be less effective for elongated, low-contrast structures. 7. Tighten speculative language in Future Work (Discussion, lines 415–423): Please distinguish clearly between planned experiments and hypotheses, avoiding language that implies expected success. ********** Figure resubmission: -->--> -->-->-->While revising your submission, we strongly recommend that you use PLOS’s NAAS tool ( https://ngplosjournals.pagemajik.ai/artanalysis ) to test your figure files. NAAS can convert your figure files to the TIFF file type and meet basic requirements (such as print size, resolution), or provide you with a report on issues that do not meet our requirements and that NAAS cannot fix.-->--> Reproducibility: -->--> -->-->To enhance the reproducibility of your results, we recommend that authors of applicable studies deposit laboratory protocols in protocols.io, where a protocol can be assigned its own identifier (DOI) such that it can be cited independently in the future. Additionally, PLOS ONE offers an option to publish peer-reviewed clinical study protocols. Read more information on sharing protocols at https://plos.org/protocols?utm_medium=editorial-email&utm_source=authorletters&utm_campaign=protocols -->--> -->-->To enhance the reproducibility of your results, we recommend that authors of applicable studies deposit laboratory protocols in protocols.io, where a protocol can be assigned its own identifier (DOI) such that it can be cited independently in the future. Additionally, PLOS ONE offers an option to publish peer-reviewed clinical study protocols. Read more information on sharing protocols at https://plos.org/protocols?utm_medium=editorial-email&utm_source=authorletters&utm_campaign=protocols -->?> PLOS Digit Health. 2026 Apr 8;5(4):e0001309. doi: 10.1371/journal.pdig.0001309.r002 Author response to Decision Letter 1 Article notes Copyright and License information Collection date 2026 Apr. PMC Copyright notice 20 Jan 2026 Attachment Submitted filename: PLOS_ResponseToReviewers.pdf pdig.0001309.s002.pdf (31.5KB, pdf) PLOS Digit Health. doi: 10.1371/journal.pdig.0001309.r003 Decision Letter 1 Ismini Lourentzou Ismini Lourentzou Section Editor Find articles by Ismini Lourentzou Author information Copyright and License information Roles Ismini Lourentzou : Section Editor © 2026 Ismini Lourentzou This is an open access article distributed under the terms of the Creative Commons Attribution License , which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. PMC Copyright notice 3 Mar 2026 Segment anything small for ultrasound: Enhancing segmentation with non-generative augmentation PDIG-D-25-00562R1 Dear Ph.D Ferreira, We are pleased to inform you that your manuscript 'Segment anything small for ultrasound: Enhancing segmentation with non-generative augmentation' has been provisionally accepted for publication in PLOS Digital Health. Before your manuscript can be formally accepted you will need to complete some formatting changes, which you will receive in a follow-up email from a member of our team. Please note that your manuscript will not be scheduled for publication until you have made the required changes, so a swift response is appreciated. IMPORTANT: The editorial review process is now complete. PLOS will only permit corrections to spelling, formatting or significant scientific errors from this point onwards. Requests for major changes, or any which affect the scientific understanding of your work, will cause delays to the publication date of your manuscript. If your institution or institutions have a press office, please notify them about your upcoming paper to help maximize its impact. If they'll be preparing press materials, please inform our press team as soon as possible -- no later than 48 hours after receiving the formal acceptance. Your manuscript will remain under strict press embargo until 2 pm Eastern Time on the date of publication. For more information, please contact [email protected]. Thank you again for supporting Open Access publishing; we are looking forward to publishing your work in PLOS Digital Health. Best regards, Ismini Lourentzou Section Editor PLOS Digital Health *********************************************************** Additional Editor Comments (if provided): Reviewer Comments (if any, and for reference): Reviewer's Responses to Questions Comments to the Author Reviewer #1: All comments have been addressed Reviewer #2: All comments have been addressed ********** publication criteria ? Is the manuscript technically sound, and do the data support the conclusions? The manuscript must describe methodologically and ethically rigorous research with conclusions that are appropriately drawn based on the data presented.? Is the manuscript technically sound, and do the data support the conclusions? The manuscript must describe methodologically and ethically rigorous research with conclusions that are appropriately drawn based on the data presented.-->?> Reviewer #1: Yes Reviewer #2: Yes ********** 3. Has the statistical analysis been performed appropriately and rigorously?-->?> Reviewer #1: Yes Reviewer #2: Yes ********** 4. Have the authors made all data underlying the findings in their manuscript fully available (please refer to the Data Availability Statement at the start of the manuscript PDF file)??> The PLOS Data policy requires authors to make all data underlying the findings described in their manuscript fully available without restriction, with rare exception. The data should be provided as part of the manuscript or its supporting information, or deposited to a public repository. For example, in addition to summary statistics, the data points behind means, medians and variance measures should be available. If there are restrictions on publicly sharing data—e.g. participant privacy or use of data from a third party—those must be specified.requires authors to make all data underlying the findings described in their manuscript fully available without restriction, with rare exception. The data should be provided as part of the manuscript or its supporting information, or deposited to a public repository. For example, in addition to summary statistics, the data points behind means, medians and variance measures should be available. If there are restrictions on publicly sharing data—e.g. participant privacy or use of data from a third party—those must be specified.--> Reviewer #1: Yes Reviewer #2: Yes ********** 5. Is the manuscript presented in an intelligible fashion and written in standard English??> Reviewer #1: Yes Reviewer #2: Yes ********** Reviewer #1: Thank you the authors for their careful and thorough responses to all points raised in the first round of review and for implementing the corresponding revisions to the manuscript. All three specific concerns I raised have been effectively addressed: 1.Figure Visualization: The updated figure (now Figure 4) uses more distinct colors, significantly improving the visual contrast between SAS and no-SAS sample distributions and making the results clearer. 2.Method Contribution Analysis: The newly added Section 3.5 (Ablation Study) and Figure 10 explicitly disentangle and demonstrate the individual and combined contributions of the scale and noise perturbations through quantitative experiments, providing a strong response to the query regarding mechanistic explanation. 3.Figure Consistency: The noise examples in Step 2 of Figure 1 have been corrected to show the four actual noise types described in the text, resolving the previous ambiguity. Upon review, the authors have sufficiently addressed and incorporated all revision suggestions from the previous round. The manuscript has seen substantial improvements in methodological explanation, experimental validation, and presentation of results. I find the revised manuscript meets the criteria for publication. Recommendation: Accept for publication. Reviewer #2: The new ablation study in Section 3.5 is a great addition, as it clearly breaks down how scale and noise perturbations each contribute to the model’s success. Using a 3% area threshold to define "small structures" is now well-supported by the empirical data you've added to Figure 2. The added transparency regarding the lower performance on large breast tumors and elongated nerves provides a more balanced view of the method's current limits. Your explanation that the black background simulates actual ultrasound zoom behavior effectively clears up concerns about domain shifts. Over all its a good manuscript. all the best. ********** Associated Data This section collects any data citations, data availability statements, or supplementary materials included in this article. Supplementary Materials S1 Fig. Graphical_abstract. (PDF) pdig.0001309.s001.pdf (1.3MB, pdf) Attachment Submitted filename: PLOS_ResponseToReviewers.pdf pdig.0001309.s002.pdf (31.5KB, pdf) Data Availability Statement The datasets used in this study are summarized in Table 1 and include both publicly available and internal datasets. Public datasets used for performance evaluation are openly accessible and are referenced in Table 1 . Additional datasets, including Kidney, Gynecology, and Follicle, are internal, non-public datasets and are not publicly available due to protected patient health information and institutional data governance restrictions. These datasets were used for model training (Kidney and Gynecology) and evaluation (Follicle). Access to these data may be considered upon reasonable request to the HealthCare Real World Data (RWD) Team at: [email protected] , subject to institutional approval and applicable data-sharing agreements. Articles from PLOS Digital Health are provided here courtesy of PLOS ACTIONS View on publisher site PDF (5.0 MB) Cite Collections Permalink PERMALINK Copy RESOURCES Similar articles Cited by other articles Links to NCBI Databases Cite Copy Download .nbib .nbib Format: AMA APA MLA NLM Add to Collections Create a new collection Add to an existing collection Name your collection * Choose a collection Unable to load your collection due to an error Please try again Add Cancel Follow NCBI NCBI on X (formerly known as Twitter) NCBI on Facebook NCBI on LinkedIn NCBI on GitHub NCBI RSS feed Connect with NLM NLM on X (formerly known as Twitter) NLM on Facebook NLM on YouTube National Library of Medicine 8600 Rockville Pike Bethesda, MD 20894 Web Policies FOIA HHS Vulnerability Disclosure Help Accessibility Careers NLM NIH HHS USA.gov Back to Top