GraphMorph centerline dice loss for small bowel volume segmentation in 3D cine-MRI

Abdominal cine-MRI provides a non-invasive approach for monitoring small bowel movement. Accurate segmentation of the small bowel is essential for measuring small bowel movement and supporting clinical applications, e.g. radiotherapy, but segmenting its volume manually on 3D cine-MRI is both challenging and time-consuming due to the large variations and changes over time in shapes, sizes, locations and poor boundary visibility of the bowel. Deep learning currently achieves promising performance in segmentation but commonly used deep learning segmentation losses, like Dice loss or Cross-Entropy, prioritize volumetric overlap, often neglecting topology. While topology-aware loss functions have improved segmentation of vascular structures such as vessels, they remain unoptimized for the small bowel due to its complex and highly variable anatomy, despite its tubular structure. In this study, we propose the GraphMorph Centerline Dice (GMC-Dice) loss for small bowel volume segmentation on 3D cine-MRI. We extract the centerlines using proposed GraphCenter method from ground truth volume, which combines the Medial Axis Transform (MAT) with anatomically informed post-processing tailored to the small bowel. To quantify how well the centerlines represent the small bowel volume, a new metric, Graph Volume Explanation (GVE), is introduced. The GraphCenter-extracted centerlines, without compromising GVE, demonstrated a superior topological representation compared to MAT or other skeletonization. Building on these centerlines, the proposed GMC-Dice loss function improved small bowel volume segmentation, achieving average Dice and centerline Dice scores of 0.696 and 0.745, respectively, outperforming existing centerline-based losses and demonstrating potential for future clinical applications.