A Novel Shape-Based Loss Function for Machine Learning-Based Seminal Organ Segmentation in Medical Imaging

Automated medical image segmentation is essential for clinical diagnosis and treatment. Standard loss functions like Dice focus on overlap but do not explicitly capture shape and morphology, often resulting in unrealistic segmentations with holes and fragmented regions. We propose a novel shape-based loss function that incorporates principal component analysis (PCA) of anatomical shape variation, enabling the network to learn underlying shape features and produce anatomically plausible segmentations. This approach discourages outlier predictions and improves segmentation validity. :contentReference[oaicite:0]{index=0}

arXiv