1
Protect the brain when treating the heart: A convolutional Neural Network for Detecting Emboli
arXiv:2604.22258v1 [cs.LG] 24 Apr 2026
Andrea Angino1,3 , Ken Trotti2 , Diego Ulisse Pizzagalli1 , Rolf Krause1,2 , Tiziano Torre4 , Stefanos Demertzis4,5
Abstract—Gaseous microemboli (GME) represent a common complication of cardiac structural interventions across both surgical and transcatheter approaches. Trans thoracic cardiac ultrasound imaging represents a convenient methodology to visualize the presence of circulating GME. However, their detection and quantification are far from trivial due to operator-dependent view, high velocity, and objects with similar structure in the background. Here, we propose an approach based on a 2.5D U-Net architecture to segment GME in space-time connected data. Such an approach yields robust detection vs. background and high segmentation accuracy while retaining real-time execution speed. These properties facilitated the integration of the proposed pipeline in patient-monitoring surgical protocols, providing the quantification of GME area over time.
I. I NTRODUCTION Cardiac structural interventions (surgical and/or transcatheter) pose significant neurological risks due to the inadvertent introduction of emboli, solid or gaseous particles, into the bloodstream. Gaseous microemboli (GME), primarily air bubbles, are especially concerning as they can travel through the circulatory system, potentially lodging in cerebral vessels and leading to silent infarctions and cognitive deficits [5]. Neurological complications arising from GME have been documented extensively, ranging from transient impairments to long-term cognitive decline in patients following cardiac procedures [11]. Existing solutions, such as transcranial Doppler ultrasound and diffusion-weighted MRI, provide a means of post-operative emboli detection, yet they lack the immediacy necessary to facilitate real-time intervention during surgery, limiting their effectiveness in mitigating intraoperative neurological risks [16]. Providing real-time feedback within the surgical workflow can therefore support the operating team by offering timely confidence estimates and improving intraoperative decision-making [12]. We focus on transthoracic cardiac echocardiography, a standardized and easily implementable modality that can be integrated into clinical practice to 1 Università della Svizzera Italiana (USI), Center for Computational Medicine in Cardiology (CCMC) 2 King Abdullah University of Science and Technology (KAUST), AMCS/CEMSE Division 3 Unidistance Suisse 4 Cardiac Surgery Department, Cardiocentro Ticino Institute, Ente Ospedaliero Cantonale (EOC) 5 Università della Svizzera Italiana (USI), Faculty of Biomedical Sciences
visualize GMEs in cardiac chambers. Despite its wide availability and suitability for intraoperative monitoring, reliable software for automatic GME detection and quantification is still lacking. This gap is largely explained by the small size of emboli, their irregular and variable appearance, and their strong similarity to surrounding tissue and common echocardiographic artifacts, which makes purely appearance-based detection unreliable. Learning-based segmentation provides a natural way to address these challenges. State-of-the-art architectures such as Fully Convolutional Networks (FCNs) [14] and Mask R-CNN [8] have been successfully applied to medical image segmentation, while U-Net remains a widely adopted reference model due to its ability to preserve fine spatial details through skip connections [17]. Beyond static segmentation, recent work has shown that U-Net-based models can also segment motile targets in grayscale images with limited resolution [3]. In the specific case of GME, however, appearance alone is often insufficient as emboli may be indistinguishable from static speckle patterns and transient artifacts in single frames, while their rapid motion across consecutive frames provides a key cue for discrimination. This motivates approaches that exploit short-term temporal context in addition to spatial features. Building on this evidence, we propose a real-time detection system based on a 2.5D U-Net architecture for GME segmentation in echocardiographic video. The model processes short sequences of consecutive frames to introduce limited temporal context while remaining computationally feasible for online deployment. By leveraging temporal continuity, the proposed approach is designed to distinguish moving emboli from the dynamic anatomical background under real-time intraoperative constraints. II. DATASET CONSTRUCTION Segmenting GME requires meticulous, timeintensive work by trained specialists capable of distinguishing microemboli from surrounding cardiac structures. Segmentation in this context presents unique challenges: while it may initially seem straightforward to delineate iperechogenic regions within each frame, the task is complicated by the fact that cardiac muscle and emboli often share similar grayscale intensities
2
resolution. Although this initial dataset is small, each frame has been subdivided into four chunks to expand the dataset, yielding a more extensive collection of localized image samples. This approach prioritizes local details, which is advantageous in this context since our goal is to identify GME features rather than capture the entire cardiac structure. The loss of global anatomical context does not pose a significant issue, as we aim to prevent overfitting to specific cardiac structures, which could be a risk given the limited number of source videos. In conclusion, we obtained a dataset with ∼ 4000 chunks, each measuring 300 × 400 pixels. Fig. 1: Color-coded overlay of two consecutive frames: the initial frame is depicted in red, while the subsequent frame appears in turquoise. The dynamic movement of GMEs is emphasized by their distinct coloring, contrasting with the white stationary elements representing the stable cardiac background.
at echocardiographic evaluation. This similarity creates ambiguity, especially near edges, where motion is the key indicator distinguishing rapidly moving emboli from the slower-moving or stationary cardiac tissue. Manual analysis typically requires moving through video frames to observe these subtle differences in movement, as GMEs are characterized by rapid positional changes compared to the relatively stable cardiac background. Figure 1 presents two overlapping subsequent frames, with the first shown in red and the second in turquoise. This color-coded overlay highlights motion cues, enabling the precise identification of emboli against the stationary cardiac background. For this study, we developed a custom annotation tool tailored to meet the specific requirements of GME segmentation in echocardiographic frames. This tool provides intuitive navigation through video frames, allowing expert annotators to segment emboli using two selection modes: a manual outline tool for precise boundary tracing and an automatic region selection tool that identifies regions based on color variation. Although the automatic mode expedites the segmentation process, it can occasionally lack precision, making manual corrections necessary. The tool also includes user-friendly features, such as an “undo” function to reverse recent actions, a “cancel” option for discarding wrongly selected regions, and an auto-save feature to ensure that all changes are securely stored. The tool can be found in [2], and can be freely downloaded and used to create datasets for similar segmentation tasks, aiming to facilitate research and data sharing in the field. The resulting dataset consists of eight echocardiographic videos, each approximately 2 seconds long, recorded at 60 frames per second in 600 × 800
III. N EURAL NETWORK STRUCTURE To provide a comparative baseline, we first evaluated classical segmentation approaches using the open-source Fiji/TrackMate 7 [7] software (Thresholding and Laplacian of Gaussian (LoG) filtering strategies). The top-left and top-right panels in Figure 2 correspond to the segmentation outputs obtained with these methods. As can be seen, both approaches capture not only the GMEs but also portions of the surrounding muscular tissue, resulting in oversegmentation. The LoG method shows improved sensitivity to GMEs but still fails to suppress responses from cardiac structures with similar intensity and texture characteristics. These limitations of traditional image processing approaches highlighted the need for a learning-based strategy capable of capturing more complex spatial and contextual features, motivating our decision to adopt a deep learning framework. A 2D U-Net architecture, which is widely recognized for its effectiveness in biomedical image segmentation [17] struggles to effectively distinguish between GME and the surrounding cardiac boundaries due to the limited size of our dataset and the intrinsic complexity of differentiating microemboli from surrounding tissue structures. Considering the importance of motion cues in differentiating emboli from background tissue, an observation made during dataset construction, the main improvement in model architecture was achieved by transitioning to a 2.5D U-Net [1]. Unlike a 3D UNet, which processes a full volumetric space-time block of data, and hence is not suitable for real-time processing, a 2.5D U-Net partially incorporates the temporal dimension by including a short sequence of frames along the time axis. This allows the network to capture essential temporal information while remaining computationally efficient for real-time applications. As shown in Figures 2.c and 2.d, the 2D model often misclassifies background resembling GMEs and underestimates the true GMEs area, whereas the 2.5D model accurately excludes the background. The choice of a U-Net architecture stems from its well-documented success in biomedical imaging tasks due to its encoder-decoder structure [18], which
3
(a) Segmentation output obtained using basic intensity (b) Segmentation result using Laplacian of Gaussian thresholding in Fiji. (LoG) filtering via TrackMate in Fiji.
(c) Preliminary segmentation results using 2D U-Net.
(d) Preliminary segmentation results using 2.5D U-Net.
Fig. 2: Preliminary segmentation comparisons using classical Fiji-based methods and deep learning models, illustrating the shift from traditional image processing to learning-based approaches.
enables detailed segmentation by leveraging contextual information. The U-Net’s skip connections allow the model to retain spatial information lost during down-sampling, crucial for accurately segmenting GME boundaries against a complex background. All patients undergoing cardiac surgery are equipped with a direct invasive arterial line, a central venous catheter, and a bladder catheter. Under conditions of general anesthesia and orotracheal intubation, a transesophageal ultrasound probe is positioned, which allows continuous monitoring of cardiac contractility, the functionality of the heart valves, and the result of the operation performed. The video frames were acquired by the echocardiographic machine (Philips Epiq) during the weaning phases from the CardioPulmonary Bypass (CPB) when any bubbles present inside the cardiac cavities are pushed forward by the resumption of pulmonary ventilation. To enable real-time processing and sharing, an HDTV video capture USB 3.0 device was used to connect the Philips Epiq to a laptop, mirroring the echocardiographic images while capturing high-definition video for NN analysis. The NN processes batches of 10 frames, introducing a delay of 0.16 seconds, during which the next batch is processed iteratively, ensuring
near real-time visualization. This setup allows doctors to perform their assessments as if the AI were not present, relying solely on the live echocardiographic feed. However, they can also refer to the laptop screen to visualize quantitative bubble estimates and the regions identified by the neural network as potential GMEs, providing an additional layer of insight. The binary segmentation procedure, described in Algorithm 1, takes a raw grayscale frame as input and applies a pretrained U-Net to produce a binary mask. The input frame is first divided into four quadrants, which are processed in batches by the network. The resulting per-pixel probabilities are reassembled into a full-frame probability map and thresholded to obtain the final binary segmentation. a) Hardware requirements: To accommodate real-time processing of a convolutional NN, we recommend a computing setup equipped with at least an 8 GB CUDA-enabled GPU and an 8-core processor. This setup ensures energy, resource efficiency, and can be physically located in the operating room near the echocardiography machine. The machine requires multiple cores so the workload can be split into several concurrent processing pipelines. One pipeline is dedicated to the NN’s inference, another
4
Algorithm 1: Binary 2.5D U-Net Segmentation Input: Frames F ∈ Rlh×600×800 , pretrained 2.5D U-Net U, threshold t ∈ [0, 1] 600×800 Output: Binary mask M ∈ {0, 1} Preprocess Partition F into patches {Fk }4k=1 ; // Fk ∈ Rlh×300×400 Form the tensor X = [F1 , F2 , F3 , F4 ] ; // X ∈ R4×lh×300×400 Network Inference P ← U(X); Aggregate probability patches into P ∈ [0, 1]600×800 ;
// P ∈ [0, 1]4×300×400
Thresholding foreach pixel Pij in P do if Pij ≥ t then Mij ← 1; else Mij ← 0;
// foreground // background
return M ;
responsible for computing the estimate by processing the segmented data through an exponential moving average, effectively prioritizing recent changes while retaining long-term trends for accuracy, and a third for generating the real-time output, which results from combining the previous contributions into a coherent visualization (See example in Figure 6). This division of tasks ensures seamless integration and uninterrupted real-time operation, even in resourceconstrained environments. A. Segmentation Performance Metrics To comprehensively evaluate segmentation performance, we consider three complementary metrics: the Frequency Bias Index (FBI), the Intersection over Union (IoU), and the Dice coefficient. Each captures different aspects of segmentation quality, allowing for a nuanced assessment of model behavior in detecting GME regions. Throughout this section, we use the following notation: TP (true positives): number of GME pixels correctly predicted as GME, • FP (false positives): number of background pixels incorrectly predicted as GME, • FN (false negatives): number of GME pixels missed by the model, • TN (true negatives): number of background pixels correctly predicted as background. •
a) Frequency Bias Index: The Frequency Bias Index (FBI) is a key performance metric in our analysis. It is commonly used in forecasting to measure the tendency of a model to over- or under-predict positive events. In our context, it is defined as the ratio between the number of pixels predicted as GME and the total number of ground-truth GME pixels, rescaled to a percentage for visualization purposes:
// Binary mask
FBI =
TP + FP × 100 = TP + FN
1+
FP − FN TP + FN
× 100.
(1) A value close to 100% indicates that the predicted segmentation has approximately the same area as the ground-truth GME, regardless of its spatial alignment. The right-hand side of Equation 1 highlights how the FBI reflects the net discrepancy between predicted and true GME regions: if FN > FP, the model underestimates the GME extent and the FBI falls below 100%. Conversely, if FP > FN, the model overestimates it and the FBI exceeds 100%. This property makes the FBI particularly suitable for our application, where capturing the full extent of the GME is critical, as a slight overestimation (Type I error dominant, FP > FN), which may lead to unnecessary deairing procedures, is more acceptable than underestimation (Type II error dominant, FN > FP), which could result in failing to intervene on patients at risk due to high GME levels. Thus, a model with an FBI slightly above 100% is preferable to one below 100%, as it reflects a conservative yet balanced estimate of the GME area. b) Intersection over Union (IoU): The Intersection over Union (IoU), or Jaccard index, quantifies the overlap between predicted and true GME regions in terms of pixel counts. It is defined as IoU =
TP × 100% TP + FP + FN
(2)
An IoU of 100% means perfect agreement, while 0% means no overlap. Because every FP and FN reduces the score equally, IoU is a balanced measure of segmentation quality. In practice, IoU values above 50% are considered acceptable, above 75% good, and values exceeding 90% indicate excellent performance in challenging medical imaging tasks.
5
c) Dice Coefficient: The Dice coefficient (also called the F1 -score for segmentation) emphasizes the harmonic mean of precision and recall. In terms of pixel counts: Dice =
2 TP × 100% 2 TP + FP + FN
(3)
Dice ranges from 0% (no overlap) to 100% (perfect overlap). By giving double weight to TP, Dice is more sensitive to small structures and less punishing of isolated FP or FN than IoU. Typical benchmarks consider Dice above 70% moderate, above 85% strong, and above 90% near perfect for fine grained segmentation. IV. N UMERICAL R ESULTS This section presents all experimental results. In Section IV-A, we investigate the effect of input length, specifically, the number of frames provided to the 2.5D U-Net. Section IV-B presents a parameter study to finalize the model configuration. Section IV-C reports the performance of the selected model. All experiments in Sections 4.1–4.3 were conducted under the same training conditions, as detailed below. All models were trained for 50 epochs with a batch size of 10, using binary cross-entropy loss and Adam optimizer. The learning rate was set to lr = 10−3 for the first 24 epochs and then reduced to lr = 10−4 thereafter, using the default momentum values of β1 = 0.9 and β2 = 0.999. Results are averaged over three runs with different random initializations. Experiments were conducted on a Raider A18 HX A9W laptop (AMD Ryzen 9 9955HX3D CPU, 64 GB RAM, NVIDIA RTX 5080 GPU with 16 GB VRAM).
Fig. 3: Training loss over 50 epochs for 2.5D UNet models with 8 base channels, varying the number of down-sampling levels (lvl ∈ {1, 2, 3, 4}) and the input history length (lh ∈ {2, 3}). All models were trained using binary cross-entropy loss and the Adam optimizer with a batch size of 10.
due to the need to wait for an additional future frame. Based on this trade-off and the limitation given by the delay, we consider two input frames sufficient to distinguish moving foreground from static background, and accordingly fix lh = 2 in all subsequent experiments. Finally, additional tests (not shown in the figure) indicate that the validation loss tends to stagnate around epoch 25, suggesting diminishing returns in model generalization despite the training loss continuing to decrease. This pattern is indicative of potential overfitting. To mitigate this effect, we stop at epoch 50 as it strikes a reasonable balance between model performance and overfitting risk. B. Trade-Off Analysis
A. Impact of Model History Length In the first experiment, we evaluate the impact of input sequence length on the performance of a 2.5D U-Net, specifically whether more than two consecutive frames offer any benefit. Due to delay constraints associated with longer histories, we limit the maximum history to three frames. Figure 3 presents the training loss curves as a function of epochs for a 2.5D U-Net for different parameters. Specifically, the base number of feature map channels is set to 8 as we vary the number of downsampling levels (lvl ∈ {1, 2, 3, 4}) and the input history length (lh ∈ {2, 3}). As expected, increasing the number of downsampling levels leads to a faster decrease in training loss. Moreover, for shallow models (lvl ∈ {1, 2}), using two frames (lh = 2) yields a small reduction in loss compare to the ones with lh = 3, while for deeper models (lvl ∈ {3, 4}), three frames (lh = 3) perform slightly better then the lh = 2 counterparts. However, the difference at epoch 50 remains below 4% and a longer history incurs extra computation and latency
The second experiment is designed to identify the 2.5D U-Net model configuration that offers the best trade-off between segmentation performance and computational cost, under the practical constraint of real-time execution on a standard laptop. Figure 4 shows the performance of the tested 2.5D U-Nets with fixed threshold t = 0.5 and lh = 2, while varying the number of base channels ch ∈ {8, 16, 32} and the number of downsampling levels lvl ∈ {1, 2, 3, 4}. The performance is measured through the classical IoU, Dice, and FBI metrics (see Section III-A), as well as average inference time per batch (right vertical axis), measured under real-time execution conditions. A single horizontal reference line guides interpretation by marking the target FBI value of 100% (indicating ideal coverage of the GME region) when read against the left axis, while simultaneously denoting the maximum acceptable inference time per batch when read against the right axis, beyond which real-time responsiveness may be compromised. All metrics are computed over the entire test set.
6
Since the FBI is the most relevant for our application. For this reason, we include error bars around the FBI values, computed as FBI±σ10 , where σ10 is the standard deviation calculated by first grouping the test dataset into batches of size 10 images and then computing the FBI over each batch. In Figure 4, as expected, we observe that increasing the number of parameters generally leads to better performance, but higher timings per batch. Thus, models with more than 100k parameters begin to achieve higher Dice and FBI values. However, beyond approximately 500k parameters, the performance gains tend to plateau. Meanwhile, the inference time per batch increases notably with model size, potentially introducing delays in real-time applications, especially on resource-constrained devices. Since the real-time application aims to estimate the area of GME regions, we prioritize models with low FBI variability and a mean FBI possibly near 100%. FBI above 100% were preferred to FBI below 100% as discussed in Section III-A, overestimating the GME area is preferable to underestimating it, as missing relevant regions may have more serious consequences. Based on this criterion, the most promising models are U-Net(ch=8, lvl=3,4), U-Net(ch=16, lvl=3), and U-Net(ch=32, lvl=4). While all four tend to slightly underestimate the GME area, probably due to the dominance of the background, the last configuration exceeds hardware constraints, making it less suitable for deployment. A final tunable parameter is the threshold t, which influences model behavior. Lowering t forces the model to accept more pixels, thereby increasing TP and FP, i.e., the numerator of FBI metric in Equation 1. Based on preliminary experiments, we set t = 0.4, which improved the performance of the selected models. Figure 5 shows the resulting metrics. As expected, FBI increases overall, while IoU and Dice remain comparable to the previous setting. We note that the U-Net(ch=8, lvl=4) and UNet(ch=16, lvl=3) models exhibit low FBI standard deviation (i.e., high reliability), moderate network size (i.e., faster inference), and strong overall performance. Both are therefore suitable for real-time applications, however, we focus on the latter, as its slightly lower FBI variance makes it the more reliable choice. This adjustment confirms that the selected model ultimately outperforms the alternatives when accounting for both predictive reliability and practical constraints. C. Model Performance Evaluation In this section, we present a comprehensive evaluation of the selected 2.5D U-Net model U(ch = 16, lvl = 3, lh = 2, t = 0.4), demonstrating its effectiveness on real data. Figure 6 shows an example of the network’s segmentation output performed in real time on a single
frame captured during surgery. At the top of the output screen, the GME area (Area b.) value is displayed, representing the estimated overall GME area. This estimate is computed by counting the number of pixels classified as positive (ones) in the network’s output mask, corresponding to the numerator of the FBI quantity. The pixel count is then rescaled to a physical measurement in mm2 , taking into account the zoom level selected by the doctor manipulating the MRI. This provides a real-time quantitative assessment of the area visible during the intervention. For visualization purposes, given that the video runs at 60 frames per second, we apply an exponential moving average (EMA) and update the GME area estimate every ten frames to ensure a readable, accurate, and interpretable value. In addition, Figure 6 reports an estimate of the processing delay, which becomes relevant when the available hardware does not fully support real-time performance; displaying this delay helps verify correct pipeline operation and allows the surgical team to assess whether latency remains within acceptable bounds. This measurement can be used during the procedure as a real-time indicator of GME presence and extent, and its reliability can be assessed directly by comparing the predicted GMEs with the live visual feedback. Moreover, in the post-operative phase, the network can be applied to provide the same quantitative information almost instantaneously, transforming the original data into a more interpretable format that may also be accessible to non-clinical personnel. V. C ONCLUSIONS This work identified the 2.5D U-Net as a practical and computationally efficient method for real-time detection of gaseous microemboli in intraoperative cardiac ultrasound. By leveraging short-term temporal context, the model distinguishes moving emboli from the cardiac background while maintaining real-time performance. The selected configuration provides the best trade-off between accuracy, GME area stability, and inference speed, enabling clinical integration. Although the study was conducted on a limited number of subjects using a single imaging device, U-Net-based architectures have shown strong adaptability across medical imaging tasks [15]. Therefore, the proposed protocol is expected to generalize to other centers with similar acquisition settings. Overall, the framework represents a step toward real-time AI-assisted monitoring of embolic load, with potential benefits for intraoperative decisionmaking and patient safety. A. Future Works Future developments will focus on extending model capabilities, improving generalization, and exploring more advanced learning paradigms to better support clinical integration.
7
Fig. 4: Segmentation performance (IoU, Dice, and FBI with standard deviation) and inference time per batch for different 2.5D U-Net architectures, evaluated with a fixed threshold of t = 0.5. Each model is identified by its configuration (channels, levels) and the corresponding number of parameters. FBI values are displayed with error bars to indicate variability across the test dataset. Inference time is plotted using a secondary y-axis.
Fig. 5: Segmentation performance (IoU, Dice, and FBI with standard deviation) and inference time per batch for different 2.5D U-Net architectures, evaluated with a threshold of t = 0.4. Each model is identified by its configuration and number of parameters.
A natural extension of the current architecture involves the adoption of transformer-based models such as TransUNet, SwinUNet, and TransBTSV2, which have demonstrated competitive or superior performance in large-scale benchmarks like ACDC and BraTS [10]. These architectures excel at capturing long-range dependencies and could prove particularly valuable in medical video segmentation tasks, where
temporal and spatial context are both critical. However, their high data requirements and sensitivity to overfitting pose challenges in low-data regimes like ours. As our dataset grows, it will become increasingly feasible to fine-tune or pre-train such models effectively for our domain-specific use case. Alternatively, unsupervised learning strategies are expected to play an increasingly central role in fu-
8
Fig. 6: Black-and-white frame acquired in real time from an echocardiogram of a patient. Red highlights indicate GMEs detected by the 2.5D Unet model U(ch = 16, lvl = 3, lh = 2, t = 0.4). overlaid in green text at the top of the image.
ture developments [9]. Manual annotation of medical videos is time-consuming and requires expert knowledge, making supervised learning difficult to scale. Approaches such as consistency regularization, pseudo-labeling, and contrastive learning offer promising avenues to leverage large amounts of unlabeled data, reducing the dependency on costly annotations while still improving generalization [4], [13]. Therefore, we believe the current 2.5D U-Net architecture remains a reliable and computationally efficient choice, well-suited to our current data constraints. It can serve as a strong foundation for future training iterations and a baseline for assessing the benefits of more complex architectures as the dataset matures. Beyond architectural considerations, future work will also focus on improving model robustness and generalization across diverse patient populations and imaging conditions. Incorporating advanced data augmentation strategies has proven effective in enhancing performance by generating plausible training examples and mitigating overfitting [6]. Additionally, integrating real-time feedback mechanisms into the clinical workflow, including the estimation of a systemic GME score, can support clinicians by providing timely confidence estimates and enhancing decisionmaking during interventions [12]. Acknowledgements: The authors acknowledge the support of Fondazione Aldo e Cele Daccò through the ’Fondo EOC-USI’. The funding was essential for carrying out the present research. R EFERENCES [1] Christoph Angermann, Markus Haltmeier, Ruth Steiger, Sergiy Pereverzyev Jr., and Elke Ruth Gizewski. Projectionbased 2.5d u-net architecture for fast volumetric segmentation. CoRR, abs/1902.00347, 2019. [2] A. Angino and K. Trotti. Aircatch. https://github.com/ AnginoA/Aircatch, 2025. GitHub repository.
[3] Paola Antonello, Diego Morone, Edisa Pirani, Mariagrazia Uguccioni, Marcus Thelen, Rolf Krause, and Diego Pizzagalli. Tracking unlabeled cancer cells imaged with low resolution in wide migration chambers via u-net class-1 probability (pseudofluorescence). Journal of Biological Engineering, 17, 01 2023. [4] Jang Hyun Cho, Utkarsh Mall, Kavita Bala, and Bharath Hariharan. Picie: Unsupervised semantic segmentation using invariance and equivariance in clustering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16794–16804, June 2021. [5] Emma M. L. Chung, Caroline Banahan, Nikil Patel, Justyna Janus, David Marshall, Mark A. Horsfield, Clément Rousseau, Jonathan Keelan, David H. Evans, and James P. Hague. Size distribution of air bubbles entering the brain during cardiac surgery. PLOS ONE, 10(4):1–11, 04 2015. [6] C. Dueholm Vestergaard, U. Elstrøm Vindelev, L. P. Muren, J. Ren, O. Nørrevang, K. Jensen, and V. T. Taasti. Data augmentation for medical imaging: A systematic literature review. Physics Imaging and Radiation Oncology, 32:100658, 2024. [7] Dmitry Ershov, Minh-Son Phan, Joanna W. Pylvänäinen, Stéphane U. Rigaud, Laure Le Blanc, Arthur Charles-Orszag, James R. W. Conway, Romain F. Laine, Nathan H. Roy, Daria Bonazzi, Guillaume Duménil, Guillaume Jacquemet, and Jean-Yves Tinevez. Trackmate 7: integrating state-of-theart segmentation algorithms into tracking pipelines. Nature Methods, 19(7):829–832, 2022. [8] Kaiming He, Georgia Gkioxari, Piotr Dollar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), October 2017. [9] Xu Ji, Joao F. Henriques, and Andrea Vedaldi. Invariant information clustering for unsupervised image classification and segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), October 2019. [10] Rabeea Fatma Khan, Byoung-Dai Lee, and Mu Sook Lee. Transformers in medical image segmentation: a narrative review. Quantitative Imaging in Medicine and Surgery, 13(12), 2023. [11] Kazuki Kihara and Kazumasa Orihashi. Investigation of air bubble properties: Relevance to prevention of coronary air embolism during cardiac surgery. Artificial Organs, 45(9):E349–E358, September 2021. © 2021 International Center for Artificial Organs and Transplantation and Wiley Periodicals, Inc. [12] A. Kurz, K. Hauser, H. A. Mehrtens, E. Krieghoff-Henning, A. Hekler, J. N. Kather, S. Fröhling, C. von Kalle, and T. J. Brinker. Uncertainty estimation in medical image classification: Systematic review. JMIR Medical Informatics, 10(8):e36427, 2022. [13] Seungho Lee, Seoungyoon Kang, and Hyunjung Shim. Selfsupervised vision transformers are efficient segmentation learners for imperfect labels, 2024. [14] Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2015. [15] Fnu Neha, Deepshikha Bhati, Deepak Kumar Shukla, Sonavi Makarand Dalvi, Nikolaos Mantzou, and Safa Shubbar. An analytics-driven review of u-net for medical image segmentation. Healthcare Analytics, 8:100416, 2025. [16] Kazumasa Orihashi and Tsuyoshi Miyata. Retained intracardiac air in cardiovascular surgery: a re-visited problem. General Thoracic and Cardiovascular Surgery, 72(7):429– 438, July 2024. Published online 2024/07/01. [17] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Nassir Navab, Joachim Hornegger, William M. Wells, and Alejandro F. Frangi, editors, Medical Image Computing and Computer-Assisted Intervention – MICCAI 2015, pages 234– 241, Cham, 2015. Springer International Publishing. [18] Nahian Siddique, Sidike Paheding, Colin P. Elkin, and Vijay Devabhaktuni. U-net and its variants for medical image segmentation: A review of theory and applications. IEEE Access, 9:82031–82057, 2021.