Semi-automatic muscle segmentation in MR images using deep registration-based label propagation

Abstract

Fully automated approaches based on convolutional neural networks have shown promising performances on muscle segmentation from magnetic resonance (MR) images, but still rely on an extensive amount of training data to achieve valuable results. Muscle segmentation for pediatric and rare diseases cohorts is therefore still often done manually. Producing dense delineations over 3D volumes remains a time-consuming and tedious task, with significant redundancy between successive slices. In this work, we propose a segmentation method relying on registration-based label propagation, which provides 3D muscle delineations from a limited number of annotated 2D slices. Based on an unsupervised deep registration scheme, our approach ensures the preservation of anatomical structures by penalizing deformation compositions that do not produce consistent segmentation from one annotated slice to another. Evaluation is performed on MR data from lower leg and shoulder joints. Results demonstrate that the proposed few-shot multi-label segmentation model outperforms state-of-the-art techniques.

Keywords: deep registration; few-shot segmentation; label propagation; musculoskeletal system.

Figure 1:

Deep registration-based label propagation from an annotated slice $\{{\mathbf{x}}_i,{\mathbf{y}}_i\}$ to the next one $\{{\mathbf{x}}_j,{\mathbf{y}}_j\}$ towards dense muscle segmentation in MR images. Please refer to the text for complete description of notations.

Figure 2:

Sample images from the two evaluated shoulder MR dataset in each plane, with corresponding labels. (Top) 3D MR scan of a pediatric shoulder of a child with obstetrical brachial plexus palsy from the MR Shoulders dataset. Four muscles of the rotator cuff are represented in color: deltoid, infraspinatus, subscapularis and supraspinatus muscles. (Bottom) 3D MR scan of a healthy adult thigh extracted from the MyoSegmenTUM dataset. Four muscle groups are represented in color: gracilis, hamstring, quadriceps femoris and sartorius. Each muscle is differentiated by leg (left-right).

Figure 3:

Averaged Dice (DSC) for each annotated slice over the MR Shoulders (a) and MyoSegmenTUM (b) datasets, in a few-shot setting. DSC is displayed with respect to the normalized axial slice number obtained by linearly scaling slice number from $\left[z_{\mathit{min}},z_{\mathit{max}}\right]$ to [0, 1] where $z_{\mathit{min}}\left(z_{\mathit{max}}\right)$ is the minimal (maximal) slice index displaying a muscle. Vertical dotted black lines represents the location of annotated slices used for training. Colored areas deal with standard deviation.

Figure 4:

Performance evolution over training epochs on MR shoulders dataset, with or without pretraining. Shown metric is the 2D DSC averaged over all subjects and muscles. Colored areas deal with standard deviation.

Figure 5:

Visual comparison (first row : axial, second : sagittal, last : coronal) on MyoSegmenTUM dataset of existing methods and our approach in the few-shot setting (Sect. 3.3).

Figure 6:

Visual comparison (coronal plane) on MR Shoulder dataset of pretraining stage influence in the few-shot setting.

Figure 7:

Visual comparison of axial slice predictions for different weighting strategies on MyoSegmenTUM dataset, in the three annotated slices setting. Images surrounded by solid lines yellow lines are annotated slices used for training and propagation.