Clinical Report: Ultrasound Segmentation Using Defined Anatomical Boundaries

Overview

This study introduces a novel approach to ultrasound segmentation by distinguishing between distinct and completed anatomical boundaries, reflecting clinical practice. Utilizing Grad-CAM, the research evaluates neural network focus areas across four ultrasound datasets and provides an open-source LEG-3D-US dataset to support reproducibility and further research.

Background

Ultrasound imaging is a widely used, non-invasive diagnostic tool valued for its safety and portability, especially in obstetrics, cardiology, and oncology. However, segmentation of ultrasound images is challenging due to noise, speckles, shadows, and low contrast, which complicate boundary delineation. Unlike CT or MRI, ultrasound relies on acoustic wave reflections, making anatomical boundary identification difficult in unclear regions. Accurate segmentation requires both anatomical and positional context, and current evaluation metrics may not fully capture a network's understanding of border information.

Data Highlights

The study evaluated three encoder-decoder neural network architectures across four ultrasound datasets: thyroid, nerve, leg, and prostate. A novel LEG-3D-US dataset was introduced, featuring complex muscle segmentation with blurred and difficult-to-delineate borders. Grad-CAM was employed to analyze network activations, focusing on distinct versus completed boundary regions to mirror clinical segmentation approaches.

Key Findings

• Segmentation networks were assessed based on their ability to delineate clear anatomical boundaries and complete unclear regions separately, aligning with clinical practice.
• Grad-CAM analysis revealed that networks prioritize different regions depending on boundary clarity, providing insight into learned representations beyond conventional accuracy metrics.
• The LEG-3D-US dataset, containing challenging muscle segmentation cases, was developed and made publicly available to facilitate reproducibility and further research.
• Existing evaluation metrics like Dice score and Hausdorff distance may not fully reflect a network's understanding of ambiguous ultrasound borders.
• Open-source datasets are critical for standardizing benchmarks and accelerating innovation in ultrasound segmentation research.

Clinical Implications

This approach enhances the interpretability of ultrasound segmentation networks by focusing on how models handle distinct versus ambiguous anatomical boundaries, which is crucial for reliable diagnosis and treatment planning. The availability of the LEG-3D-US dataset encourages broader validation and development of robust segmentation tools applicable to complex clinical scenarios. Clinicians and researchers should consider both boundary delineation and completion capabilities when evaluating segmentation algorithms.

Conclusion

Distinguishing between distinct and completed anatomical boundaries in ultrasound segmentation provides a more clinically relevant evaluation of neural networks. The integration of Grad-CAM and open-source datasets like LEG-3D-US supports improved understanding and development of segmentation models tailored to ultrasound's unique challenges.

References

1. Litjens et al. 2017 -- A survey on deep learning in medical image analysis
2. Ronneberger et al. 2015 -- U-Net: Convolutional Networks for Biomedical Image Segmentation
3. Selvaraju et al. 2017 -- Grad-CAM: Visual Explanations from Deep Networks via Gradient-based Localization
4. Zhou et al. 2016 -- Learning Deep Features for Discriminative Localization
5. Open-source ultrasound datasets -- Various sources