Training-only ultrasound-specific augmentation for ovarian tumor segmentation across B-mode and contrast-enhanced...

Ovarian tumor segmentation models trained on conventional ultrasound may be sensitive to contrast-enhanced ultrasound appearance. We evaluated whether training-only combined ultrasound-specific augmentation could improve robustness without adding inference-time complexity.

We used MMOTU image-mask pairs, including 820 development B-mode ultrasound images, 382 internal two-dimensional ultrasound images, and 170 contrast-enhanced ultrasound images as a CEUS domain-shift stress test. A lightweight Residual U-Net was compared with the same backbone trained using photometric, blur, and low-amplitude noise augmentation. Performance was evaluated using overlap, boundary, and pixel-calibration metrics. Paired CEUS differences were tested with Wilcoxon signed-rank tests, and effect sizes, 95% confidence intervals, and case-level failure rates were reported.

The combined augmentation preserved internal performance, with Dice of 0.745 (0.196) compared with 0.750 (0.191) for Residual U-Net. On contrast-enhanced ultrasound, it improved Dice from 0.468 (0.190) to 0.476 (0.192), intersection over union from 0.326 (0.167) to 0.334 (0.171), Boundary F1 from 0.037 (0.035) to 0.041 (0.032), and pixel expected calibration error from 0.314 (0.143) to 0.297 (0.138). The mean paired CEUS Dice improvement was 0.008 (95% CI -0.0002 to 0.0167); 62.9% of cases improved and 37.1% worsened. However, the proportion of CEUS cases with Dice below 0.5 decreased from 55.9 to 51.8%, indicating that CEUS segmentation remained suboptimal.

Training-only combined ultrasound-specific augmentation yielded a small but directionally favorable CEUS improvement without added inference-time complexity. These findings support training-distribution design as a practical low-cost strategy for modality-shift robustness and warrant further patient-level and multi-center validation.