Quantification of Inter-Muscular Adipose Infiltration in Calf/Thigh MRI using Fully and Weakly Supervised Semantic Segmentation

Background: Infiltration of subcutaneous-adipose-tissue (SAT) is a clinical manifestation of muscular-dystrophies. Quantification of fat infiltration based on MRI techniques is an accurate marker of disease severity. In order to provide physicians with a precise biomarker, a prerequisite of reliable segmentation of muscle-tissue and inter-muscular-adipose-tissue (IMAT) is needed.

Methods: we calculated the ratio of the area of IMAT to whole muscle-region using two stages: 1.Muscle-region segmentation: we employed U-net architecture for the segmentation of muscle-region. The SAT, bone and bone marrow were excluded. The Dice-loss was optimized during training. 2.IMAT and Healthy-muscle classification: Following segmentation, weakly-supervised pixel classification is performed to differentiate between diseased muscle pixels from healthy ones. We implemented a patch-based-deep-convolutional-auto-encoder with a triple-loss constraint to learn a semantic-feature-representation and applied k-means in the embedded-space to classify the pixels into two clusters. The loss that was optimized consists of two parts: 1.Reconstruction-loss: mean-squared-error, and 2.Triplet-loss on triplet patches. 14 axial MR-scans of patients’ thigh/calf suffering from Dysferlinopathy were used for training

Results: The test set includes three patients with mild, moderate and severe fat infiltration. We achieved a Dice of 0.96, 0.91 and 0.93 for muscle-region-segmentation, healthy-muscle-segmentation and IMAT, respectively. The clustering performance was evaluated with the normalized-mutual-information, accuracy of clustering, and adjusted-Rand-index and achieved 0.63, 0.93 and 0.75, respectively

Conclusions: We demonstrated the robustness of our method to segment and classify muscle tissue in the presence of MRI artifacts. Excellent performance was demonstrated on moderate and severe cases of fat-infiltration, where other conventional methods fail.