Stool Segmentation Using Anatomical Multi-task Learning and Multi-channel Input for Constipation Diagnosis

Abstract

Chronic constipation is a common gastrointestinal disorder affecting approximately 15% of the world's population and significantly degrading quality of life. Accurate segmentation of stool regions in abdominal X-rays is essential for diagnosis, yet it remains a challenging task even for specialists due to low contrast and indistinct boundaries. This study proposes an automated stool segmentation method that utilizes gas information as an anatomical guide. Using 282 expert-annotated abdominal X-ray images, we compared the performance of four U-Net-based architectures: (1) a Baseline single-task U-Net; (2) MT-UNet (Multi-Task U-Net), which learns gas segmentation as an auxiliary task; (3) MC-UNet (Multi-Channel U-Net), which uses a gas region mask as an additional input channel; and (4) MTMC-UNet, a novel integration of the two. Evaluation via 5-fold cross-validation demonstrated that the MTMC-UNet achieved the highest Dice coefficient (0.732), a significant improvement over the Baseline (0.710). Notably, Recall substantially increased from 0.675 to 0.735, effectively reducing missed detections of clinically important stool regions. For the prediction accuracy of the newly introduced Stool Volume Score (SVS), the MC-UNet showed the highest correlation coefficient (r = 0.859). This approach enables even less experienced clinicians to perform reliable stool volume evaluation from X-rays, contributing to the standardization and objectivity of constipation diagnosis.