Deep convolutional encoder-decoders for deltoid segmentation using healthy versus pathological learning transferability

Abstract

Shoulder muscle segmentation in patients with obstetrical brachial plexus palsy is a challenging task from magnetic resonance images. A reliable fully-automated method could greatly help clinicians to plan therapeutic interventions. Among various structures, shoulders comprise a rounded and triangular-shaped muscle located on top: the deltoid. The purpose of this work consists in investigating the feasibility of pathological deltoid segmentation using deep convolutional encoder-decoders. Given a limited amount of available annotated images, we study learning transferability from healthy to pathological data by comparing different learning schemes in terms of model generalizability. Extended versions of convolutional encoder-decoder architectures using an encoder pre-trained on non-medical data are proposed to improve the delineation accuracy. Promising results obtained on a dataset of 24 shoulder examinations offer new insights for force inference in musculo-skeletal disorder management.