Objective:

To evaluate whether training-only combined ultrasound-specific augmentation can improve robustness of ovarian tumor segmentation models without adding inference-time complexity.

Approach:

• Data Used: Utilized MMOTU image-mask pairs, including 820 B-mode ultrasound images, 382 internal two-dimensional ultrasound images, and 170 contrast-enhanced ultrasound images.
• Model Comparison: Compared a lightweight Residual U-Net with a version trained using photometric, blur, and low-amplitude noise augmentation.
• Performance Evaluation: Assessed performance using overlap, boundary, and pixel-calibration metrics; applied Wilcoxon signed-rank tests for paired CEUS differences.

Key Findings:

• Thecombinedaugmentationpreservedinternalperformance,withaDicescoreof0.745(0.196)comparedwith0.750(0.191)forResidualU-Net.Oncontrast-enhancedultrasound,itimprovedDicefrom0.468(0.190)to0.476(0.192),intersectionoverunionfrom0.326(0.167)to0.334(0.171),BoundaryF1from0.037(0.035)to0.041(0.032),andpixelexpectedcalibrationerrorfrom0.314(0.143)to0.297(0.138).ThemeanpairedCEUSDiceimprovementwas0.008(95%CI-0.0002to0.0167);62.9%ofcasesimprovedand37.1%worsened.TheproportionofCEUScaseswithDicebelow0.5decreasedfrom55.9%to51.8%.

Interpretation:

Training-only combined ultrasound-specific augmentation yielded a small but directionally favorable improvement in CEUS segmentation without added inference-time complexity.

Limitations:

• CEUS segmentation remained suboptimal despite improvements.
• Findings require further validation at the patient level and across multiple centers.

Conclusion:

The study supports training-distribution design as a practical low-cost strategy for modality-shift robustness.