Clinical Scorecard: Analysis of Ultrasound Segmentation Utilizing Unique and Defined Anatomical Boundaries

At a Glance

Key Highlights

• Novel approach distinguishing distinct and completed boundaries in ultrasound segmentation mirroring medical practitioners' evaluation techniques
• Use of Grad-CAM to highlight regions prioritized by segmentation networks across multiple ultrasound datasets
• Release of LEG-3D-US dataset with segmentation at different depth levels including interpolated areas to facilitate reproducibility and further research

Guideline-Based Recommendations

Diagnosis

• Utilize ultrasound segmentation to aid diagnosis by analyzing organ-specific regions with attention to anatomical boundaries
• Consider both clear and unclear boundary regions separately for accurate interpretation

Management

• Incorporate neural network segmentation methods evaluated by border-focused metrics rather than solely prediction accuracy
• Apply Grad-CAM or similar interpretability tools to understand network decision-making in ultrasound segmentation

Monitoring & Follow-up

• Continuously assess segmentation network performance on distinct versus completed borders to ensure robustness
• Use open-source datasets like LEG-3D-US for benchmarking and validation

Risks

• Be aware of challenges posed by ultrasound artifacts such as noise, speckles, shadows, and low contrast affecting segmentation accuracy
• Recognize that traditional metrics (Dice score, Hausdorff distance) may not fully capture border delineation quality

Patient & Prescribing Data

Patients undergoing ultrasound imaging for various organs including thyroid, nerve, leg muscles, prostate

Accurate segmentation supports diagnosis and treatment planning by improving localization of structures and lesions; interpretability methods enhance trust in automated segmentation

Clinical Best Practices

• Segment ultrasound images by first delineating distinct anatomical boundaries, then completing unclear regions using prior anatomical knowledge
• Evaluate segmentation models not only by overlap metrics but also by their learned representation of border regions using interpretability tools like Grad-CAM
• Utilize and contribute to open-source ultrasound datasets to standardize evaluation and foster collaboration
• Consider anatomical and positional context in dynamic ultrasound imaging for accurate structure localization

References

• Ultrasound imaging advantages and applications
• Challenges in ultrasound image segmentation
• Medical practitioners’ approach to ultrasound segmentation
• Grad-CAM: Visual explanations from deep networks
• Interpretability methods in medical imaging
• Open-source ultrasound datasets
• LEG-3D-US dataset release