Deep Learning Approaches for Automated Classification of Muscular Dystrophies from MRI

The diagnosis of specific types of muscular dystrophies (MD) is mainly done through genetic testing. As this does not always provide an unambiguous result, muscle MRI images are often examined to confirm or verify the diagnosis as each MD type affects the muscles in a specific pattern. Different deep learning approaches (ResNet50 model pretrained on RadImageNet, auto-encoder model trained from scratch, segmentation U-Net model trained for muscle segmentation) were investigated to obtain image features from Dixon MRI of the proximal leg that were used for discriminating between cases with Becker Muscular Dystrophy (n = 18), Limb-Girdle Muscular Dystrophy R12 (n = 13) or no MD (n = 16). The results are compared with classification by a conventional random forest (RF) classifier using the fat fraction percentage per muscle as features. The RF classifier and the segmentation U-Net deep learning approach performed best with an average AUC of 0.957 and 0.934 respectively. Local interpretable model-agnostic explanations (LIME) were used to explain the decisions of the RF model. Different fat replacement patterns for BMD and LGMDR12 observed in the glutei, adductors and vasti as described in literature were in part confirmed by the explanations.