A clinically informed automated evaluation pipeline for medical image segmentation based on Medical Similarity Index

Accurate tissue delineation is essential in radiotherapy; however, conventional segmentation metrics mainly quantify geometric overlap and lack clinical interpretability. We proposed an automated Python-based evaluation pipeline using a bidirectional local distance-based metric that pairs test and reference contour points after center-of-mass correction and computes a similarity score from averaged Euclidean distances. The framework supports multislice images, multiple masks per slice, and concave mask separation, with open-source code provided. The method was demonstrated on fibroid and prostate MRI datasets, using 233 training cases and 12 test cases. In test examples, overlap scores exceeded 0.90, while Medical Similarity Index scores decreased to approximately 0.40.