Muscle segmentation in MRI volumes of lower limbs: Centralized or Federated learning?

The purpose of this work is to segment muscles in MRI volumes of lower limbs, comparing the performances of two different techniques: a 3D neural network, trained locally on a manually segmented dataset, and DAFNE (Deep Anatomical Federated Network), an open-source software trained in a decentralized collaborative manner using federated learning. Federated learning is a novel technique that involves different clients that collaborate without sharing the data. Single models are trained locally by the centers. These models are then aggregated together on the server side and the updated model is shared with each center. The dataset has been granted by Radbound University Medical Center and contains 166 lower limb volumes from patients with two different pathologies, Fascioscapulohumeral muscular Distrophy (FSHD) and Myotonic Dystrophy Type 1 (MD1), and healthy volunteers acquired using the Dixon technique. It includes 82 volumes of thigh and 84 volumes of leg, both monolateral and bilateral. For MD1 group there can be from one to three different volumes of the same patient, while for the other two groups only one volume per patient is provided. Two kinds of division of the dataset into training, validation and test set have been tested: the first one randomly extracts volumes according to set specific percentages from each type of pathological and healthy group, while the second one randomly divides the pathological volumes into training and validation set and assigns all the healthy volumes to test set. At the end of each division the training set consists of aproximatly 60 volumes and the validation and test set of about 10 volumes each. Thigh and leg volumes have been considered separately, so two 3D neural networks of the SwinUNETR type, one for the leg, that segments 9 different muscles, and one for the thigh, that segments 12 different muscles, have been trained using the MONAI library, with a pipeline designed specifically for the dataset. Five different loss functions have been tested and the results have been compared using DICE and Haussdorff Distance at the 95° percentile (HD95). Despite the unquestionable potential of DAFNE, the locally trained 3D Neural Networks, both for thigh and leg, allowed to obtain higher values of DICE and lower values of Haussdorff Distance than the segmentations given by DAFNE. In particular, for the segmentations of the volumes in test set with the 3D neural network for thigh it can be achieved a DICE of 0.7608+-0.04629 and a HD95 of 6.0260+-1.4381, while the neural network for leg allows to obtain a DICE of 0.7176+-0.0267 and a HD95 of 6.4740+-0.5157. According to the performances obtained in this work, even if with DAFNE there are great advantagies in therms of data protection and the way in which the available model can be updated and shared, at the moment a locally trained 3D neural network remains the most suitable choice for the provided dataset.