Unified Review and Benchmark of Deep Segmentation Architectures for Cardiac Ultrasound on CAMUS

TL;DR

This work addresses the lack of comparability across cardiac ultrasound segmentation studies by delivering a unified, reproducible benchmark for three prominent architectures—U‑Net, Attention U‑Net, and TransUNet—on CAMUS under harmonized preprocessing and evaluation. It systematically analyzes the effects of data representation (NIfTI vs PNG), data pairing strictness, and self‑supervised pretraining, and introduces a SAM‑based pseudo‑labeling framework to leverage unlabeled frames. The findings show that native NIfTI data yields the highest Dice scores (~94%), while 16‑bit PNGs cause modest losses; Attention U‑Net improves boundary accuracy in challenging regions, and TransUNet delivers the best global shape consistency and LA segmentation, especially with SSL. The study also demonstrates that SAM masks can meaningfully augment training data and improve generalization, and it provides practical guidance for preprocessing fidelity, architecture choice, and semi/fully‑supervised strategies. Collectively, the work offers a rigorous, reusable benchmark framework and points toward scalable self‑supervision and GPT‑assisted annotation pipelines to accelerate data curation and clinical deployment in cardiac ultrasound segmentation.

Abstract

Several review papers summarize cardiac imaging and DL advances, few works connect this overview to a unified and reproducible experimental benchmark. In this study, we combine a focused review of cardiac ultrasound segmentation literature with a controlled comparison of three influential architectures, U-Net, Attention U-Net, and TransUNet, on the Cardiac Acquisitions for Multi-Structure Ultrasound Segmentation (CAMUS) echocardiography dataset. Our benchmark spans multiple preprocessing routes, including native NIfTI volumes, 16-bit PNG exports, GPT-assisted polygon-based pseudo-labels, and self-supervised pretraining (SSL) on thousands of unlabeled cine frames. Using identical training splits, losses, and evaluation criteria, a plain U-Net achieved a 94% mean Dice when trained directly on NIfTI data (preserving native dynamic range), while the PNG-16-bit workflow reached 91% under similar conditions. Attention U-Net provided modest improvements on small or low-contrast regions, reducing boundary leakage, whereas TransUNet demonstrated the strongest generalization on challenging frames due to its ability to model global spatial context, particularly when initialized with SSL. Pseudo-labeling expanded the training set and improved robustness after confidence filtering. Overall, our contributions are threefold: a harmonized, apples-to-apples benchmark of U-Net, Attention U-Net, and TransUNet under standardized CAMUS preprocessing and evaluation; practical guidance on maintaining intensity fidelity, resolution consistency, and alignment when preparing ultrasound data; and an outlook on scalable self-supervision and emerging multimodal GPT-based annotation pipelines for rapid labeling, quality assurance, and targeted dataset curation.

Figures (7)

• Figure 1: This figure presents a comprehensive overview of CVDs wang2024artificial, categorized by the primary anatomical structures they affect and their corresponding physiological functions. It also outlines the key medical imaging modalities used for CVD diagnosis and analysis, emphasizing the growing contribution of AI in enhancing image-based assessment. The top section illustrates examples of non-vascular anatomical structures, including (a) ventricles leclerc2019deep (b) atria li2022medical, and (c) myocardium ye2021deeptag.
• Figure 4: Unified workflow for cardiac ultrasound segmentation combining literature review, strict preprocessing, model training, and evaluation. The left block summarizes the survey of cardiac imaging literature and datasets (including CAMUS and ACDC). The preprocessing stage applies PNG-16-bit conversion, strict manifest pairing, normalization, and quality checks. The modeling stage integrates U-Net, Attention U-Net, and TransUNet architectures, supported by SimCLR pretraining and SAM-based pseudo-labeling to improve encoder initialization and make use of unlabeled frames. The evaluation module compares model performance across architectures, quantifies gains from SSL, and highlights improvements in segmentation accuracy and anatomical consistency. The workflow emphasizes how preprocessing fidelity, SAM, and SSL improve reproducibility and strengthen downstream segmentation outcomes.
• Figure 5: Pipeline for SAM-based pseudo-label generation and integration into the semi-supervised training framework. The process begins with echocardiographic frames in PNG-16 bit format, which are passed through SAM in automatic mode to produce raw segmentation masks. These masks are stored as JSON and PNG files, filtered by predicted IoU ($\geq 0.7$), minimum area ($\geq 200$ pixels), and top 3 mask selection. The remaining masks are merged into structured label maps with three cardiac regions (endocardium, myocardium, and atrium) and incorporated into a semi-supervised training setup using GT and SAM pseudo-labels weighted at 1.0 and 0.5, respectively.
• Figure 6: Architectural diagram of the original U-Net model ronneberger2015u. The network follows a symmetric encoder–decoder structure with skip connections that copy and crop feature maps from the contracting path to the expanding path, preserving spatial detail. Each encoder stage consists of two $3\times3$ convolution layers with ReLU activation followed by $2\times2$ max-pooling for downsampling. The decoder mirrors this structure using $2\times2$ up-convolutions to progressively recover spatial resolution. The final $1\times1$ convolution maps feature channels to the segmentation output. This architecture enables precise localization and has become the foundational backbone for biomedical image segmentation.
• Figure 7: Architecture of the Attention U-Net model. The network extends the classical U-Net by inserting Attention Gates (red) into each skip connection. These gates receive a gating signal from the decoder and filter encoder features before concatenation, enabling the model to emphasize relevant cardiac regions (e.g., myocardium and atrial walls) and suppress background noise. Convolution, pooling, upsampling, and concatenation operations follow the standard U-Net formulation.