Abstract
Muscular dystrophies (MD) are a class of rare genetic diseases resulting in progressive muscle weakness affecting specific muscle groups, depending on the type of disease. Disease progression is characterized by the gradual replacement of muscle tissue by fat, which can be assessed with fat-sensitive magnetic resonance imaging (MRI) and objectively evaluated by quantifying the fat fraction percentage (FF%) per muscle. Volumetric quantification of fat replacement over the full 3D extent of each muscle is more precise and potentially more sensitive than 2D quantification in few selected slices only, but it requires an accurate 3D segmentation of each muscle individually, which is time consuming when this has to be performed manually for a large number of muscles. A reliable, largely automated approach for 3D muscle segmentation is thus needed to facilitate the adoption of fat fraction quantification as a measure of MD disease progression in clinical routine practice, but this is challenging due to the variable appearance of the images and the ambiguity in the discrimination of the contours of adjacent muscles, especially when the normal image contrast is affected and diminished by the fat replacement. To deal with these challenges, we used deep learning to train AI-models to segment the muscles in the proximal leg from knee to hip in Dixon MRI images of healthy subjects as well as patients with MD. We demonstrate state-of-the-art segmentation results of all 18 muscles individually in terms of overlap (Dice score, DSC) with the manual ground truth delineation for images of cases with low fat infiltration (mean overall FF%: 11.3%; mean DSC: 95.3% per image, 84.4-97.3% per muscle) as well as with medium and high fat infiltration (mean overall FF%: 44.3%; mean DSC: 89.0% per image, 70.8-94.5% per muscle). In addition, we demonstrate that the segmentation performance is largely invariant to the field of view of the MRI scan, is generalizable to patients with different types of MD and that the manual delineation effort to create the training set can be drastically reduced without significant loss of segmentation quality by delineating only a subset of the slices.