Auto Segmentation of Lower Limb Calf Muscles from Diffusion-Weighted Magnetic Resonance Images using Deep Learning

This study aimed to automate the segmentation of lower limb calf muscles from diffusion-weighted MR images using deep learning, towards the quantification of microvascular perfusion and diffusion beyond traditional ankle brachial index metric. Using manually annotated regions of interest from 24 healthy volunteers as ground truth, this study trained 2D U-Net and conditional GAN models, achieving an average Dice coefficient score of 0.7 for segmentation accuracy. Despite challenges like limited dataset size and model capacity tuning, the approach demonstrated promising results. The developed autosegmentation model from this study will enable future faster, more standardized, and less subjective analysis of DWI images, with future plans to further improve model robustness and to expand the number of acquired datasets to include more PAD patients for training and testing.