3655 - Foreground Subtraction Augmentation: Bridging the Gap between Contrast and Non-Contrast CT Segmentation

Purpose/Objective(s): Organ segmentation in CT imaging is crucial for diagnosis, treatment planning, and automated contouring. Models trained on contrast-enhanced (CON) CT scans often perform poorly on non-contrast (NONCON) scans due to differences in Hounsfield Unit (HU) distributions. This study introduces Foreground Subtraction Augmentation (FSA), which reduces the HU values of organ regions in CON images by subtracting organ mask areas with known HU differences per organ, making them resemble NONCON images. Integrating FSA-generated images into training aims to improve segmentation accuracy on NONCON scans.

Materials/Methods: A dataset of 191 CON training samples and 30 NONCON validation samples was used. Two setups were tested: a baseline model trained on unaltered CON images and an FSA-augmented model incorporating HU-adjusted images. The nnUnet 3D base model was employed with weighted Dice + CE loss, deep supervision, and an Adam optimizer (initial learning rate 1e-4, decayed to 1e-7 over 500 epochs with early stopping).

Results: FSA significantly improved segmentation accuracy on NONCON scans compared to baseline training. The Dice Similarity Coefficient (DSC) increased for multiple organs: Key improvements include Gallbladder (0.5270 to 0.6214), Kidney_L (0.8704 to 0.8946), and Liver (0.9243 to 0.9448), all of which showed statistically significant differences. The overall average DSC increased from 0.7570 to 0.7838 with significant statistical difference, reinforcing the effectiveness of FSA in mitigating contrast-related segmentation challenges.

Conclusion: Foreground Subtraction Augmentation effectively improves segmentation performance on NONCON scans without requiring additional NONCON training data. By simply modifying organ HU values through subtraction, FSA helps bridge the domain gap between CON and NONCON images. Future research will explore further refinements and applications across broader anatomical structures and imaging modalities.