Impact of MR sequences choice on deep learning segmentation of muscles

Medical image segmentation is a critical step for many medical studies. We address the problem of muscle segmentation on MRI images using Dixon sequences and explore the impact on the segmentation results when combining the four Dixon sequences available. Different combinations were put to test using two UNet-based architectures. One used an early fusion and input the images in the same encoder, while the other used late fusion, which learns the features from the images in separated encoders and then concatenates and decodes them as a whole. Our results show that the T1 water-only image is the most appropriate image for muscle segmentation in our database and that both early and late fusion approaches did not yield significantly different results. Thus, appropriate check of most adequate contrast to consider is feasible and recommended to exquisitely match to the observed population and the early fusion architecture appears to be the most efficient design to do so when dealing with such muscle segmentation task.