Clinical Report: Ultrasound-Specific Augmentation for Enhanced Segmentation of Ovarian Tumors
Overview
This study evaluates the effectiveness of ultrasound-specific augmentation in improving segmentation of ovarian tumors in both B-mode and contrast-enhanced ultrasound images. The results show a slight improvement in segmentation metrics.
Background
Ovarian cancer is a leading cause of gynecologic cancer mortality, making accurate imaging and segmentation critical for diagnosis and treatment planning. Ultrasound is commonly used for assessing adnexal masses, yet segmentation challenges persist due to variations in imaging modalities.
Data Highlights
Metric
Combined Augmentation
Residual U-Net
Dice Score
0.476
0.468
Intersection over Union
0.334
0.326
Boundary F1
0.041
0.037
Pixel Expected Calibration Error
0.297
0.314
Key Findings
Combined augmentation preserved internal performance with a Dice score of 0.745.
Contrast-enhanced ultrasound segmentation improved Dice from 0.468 to 0.476.
62.9% of cases showed improvement in segmentation metrics with combined augmentation.
The proportion of CEUS cases with Dice below 0.5 decreased from 55.9% to 51.8%.
Training-only combined augmentation did not add inference-time complexity.
Clinical Implications
The findings indicate that training models with ultrasound-specific augmentation can enhance segmentation performance in CEUS without complicating the inference process.
Conclusion
The study demonstrates that ultrasound-specific augmentation can yield modest improvements in segmentation performance for ovarian tumors in CEUS.
