299PA generalizable deep-learning muscle segmentation model for multicentre and multi-study muscle MRI in neuromuscular diseases

Abstract

Neuromuscular diseases (NMDs) are a heterogeneous group of rare conditions that impair muscle and nerve function, leading to progressive weakness and disability. Intramuscular replacement of muscle by fat, measured through muscle MRI (MRI), is a robust imaging biomarker of disease severity and progression. Its accurate quantification requires precise segmentation of individual muscles, a process that is traditionally manual, time-consuming, and prone to inter-operator variability. Recent deep learning (DL) approaches have demonstrated the potential to automate muscle segmentation. However, most published models use data from a single site or protocol, with their performance often degrading when applied to new environments. This limitation hinders the clinical adoption and scalability of these tools in real-world, multicenter settings. To address this gap, we developed a DL-based tool for automatic segmentation of individual muscles in T1-weighted MRI of the pelvis and lower limbs. The system comprises three convolutional neural networks tailored to a specific anatomical region (pelvis, thighs or lower legs) and trained on a dataset of 253 scans of 11 distinct NMDs and undiagnosed cases, from 14 sites. External validation was incorporated during training and evaluation to explicitly assess and promote generalizability. For every anatomical region, the resulting automatic segmentation models achieved high accuracy (F1-score pelvis: 09626, thighs: 0.9624 and lower leg: 0.9682) and efficiency (segmentation time under one minute), even on scans of varying resolution, protocol and disease severity. The model showed lower performance on muscles oriented in parallel to the axial plane, such as the internal and external obturators and quadratus femoris. In conclusion, our study presents a scalable, accurate, and generalizable automated muscle segmentation tool for skeletal muscle MRIs of the pelvis and lower limbs. By leveraging a heterogeneous, multi-site dataset and explicitly incorporating external validation, we demonstrate that high performance can be maintained across diverse clinical environments and imaging conditions. This advancement holds significant promise for streamlining clinical workflows, enhancing disease monitoring, and accelerating large-scale neuromuscular research efforts.