Healthy versus pathological learning transferability in shoulder muscle MRI segmentation using deep convolutional encoder-decoders

Abstract

Fully-automated segmentation of pathological shoulder muscles in patients with musculo-skeletal diseases is a challenging task due to the huge variability in muscle shape, size, location, texture and injury. A reliable automatic segmentation method from magnetic resonance images could greatly help clinicians to diagnose pathologies, plan therapeutic interventions and predict interventional outcomes while eliminating time consuming manual segmentation. The purpose of this work is three-fold. First, we investigate the feasibility of automatic pathological shoulder muscle segmentation using deep learning techniques, given a very limited amount of available annotated pediatric data. Second, we address the learning transferability from healthy to pathological data by comparing different learning schemes in terms of model generalizability. Third, extended versions of deep convolutional encoder-decoder architectures using encoders pre-trained on non-medical data are proposed to improve the segmentation accuracy. Methodological aspects are evaluated in a leave-one-out fashion on a dataset of 24 shoulder examinations from patients with unilateral obstetrical brachial plexus palsy and focus on 4 rotator cuff muscles (deltoid, infraspinatus, supraspinatus and subscapularis). The most accurate segmentation model is partially pre-trained on the large-scale ImageNet dataset and jointly exploits inter-patient healthy and pathological annotated data. Its performance reaches Dice scores of 82.4%, 82.0%, 71.0% and 82.8% for deltoid, infraspinatus, supraspinatus and subscapularis muscles. Absolute surface estimation errors are all below 83 mm² except for supraspinatus with 134.6 mm². The contributions of our work offer new avenues for inferring force from muscle volume in the context of musculo-skeletal disorder management.

Keywords: Deep convolutional encoder-decoders; Healthy versus pathological transferability; Musculo-skeletal disorders; Obstetrical brachial plexus palsy; Shoulder muscle segmentation.

Fig. 1.

Groundtruth segmentation of pathological shoulder muscles including deltoid as well as infraspinatus, supraspinatus and subscapularis from the rotator cuff. Axial, coronal and sagittal slices are extracted from a 3D MR examination acquired for a child with obstetrical brachial plexus palsy.

Fig. 2.

Three different learning schemes (P, HP, A) involved in a leave-one-out setting for deep learning-based pathological shoulder muscle segmentation.

Fig. 3.

Extension of U-Net (Ronneberger et al., 2015) by exploiting as encoder a slightly modified VGG-16 (Simonyan and Zisserman, 2014) with weights pre-trained on ImageNet (Russakovsky et al., 2015), following (Iglovikov and Shvets, 2018; Iglovikov et al., 2018). The decoder is modified to get an exactly symmetrical construction while keeping skip connections.

Fig. 4.

Box plots on Dice, Jaccard, Cohen’s kappa and absolute surface error (ASE) slice-wise scores over the pathological dataset using convolutional encoder-decoders (U-Net (Ronneberger et al., 2015), v16U-Net, v16pU-Net) embedded with learning schemes P, HP and A. Dashed green and solid orange lines respectively represent means and medians.

Fig. 5.

Deltoid segmentation accuracy using U-Net (Ronneberger et al., 2015) with learning schemes P, HP and A for each annotated slice of the whole pathological dataset. Top raw shows Dice scores (%) with respect to the normalized axial slice number obtained by linearly scaling slice number from [z_min, z_max] to [0, 1] where {z_min, z_max} are the minimal and maximal axial slice indices displaying the deltoid. Bottom row displays concordance between groundtruth and predicted deltoid muscle surfaces in mm². Black line indicates perfect concordance.

Fig. 6.

Automatic pathological deltoid segmentation using U-Net (Ronneberger et al., 2015) embedded with learning schemes P, HP and A. Groundtruth and estimated delineations are in green and red respectively. Displayed results cover the whole muscle spatial extent for L-P-0103 examination.

Fig. 7.

Automatic pathological segmentation of infraspinatus, supraspinatus and subscapularis using U-Net (Ronneberger et al., 2015) with training on both healthy and pathological data simultaneously (A). Groundtruth and estimated delineations are in green and red respectively. Displayed results cover the whole muscle spatial extents for R-P-0447 (top), R-P-0660 (middle) and R-P-0134 (bottom) examinations.

Fig. 8.

Deltoid segmentation accuracy using U-Net (Ronneberger et al., 2015), v16U-Net and v16pU-Net with learning scheme A for each annotated slice of the whole pathological dataset. Top raw shows Dice scores with respect to normalized axial slice number. Bottom row displays concordance between groundtruth and predicted deltoid surfaces. Black line indicates perfect concordance.

Fig. 9.

Automatic pathological segmentation of deltoid, infraspinatus, supraspinatus and subscapularis using U-Net (Ronneberger et al., 2015), v16U-Net and v16pU-Net with training on both healthy and pathological data simultaneously (A). Groundtruth and estimated delineations are in green and red respectively. 8 pathological examinations among the 12 available are involved to provide valuable insight into the overall performance.