Deep Learning based Fully Automatic Quantification of Rotator Cuff Tears from MRI

Rotator cuff tears are the most common source of shoulder pain. Many factors can be considered to choose the right surgical treatment procedure. The most important factors are tear thickness (partial vs. full), tear size, tear shape, and muscle quality. The aim of this work was the fully automated quantification and classification of a full-thickness posterosuperior rotator cuff tear from MR images using a deep learning based approach. A complete new approach to automatically quantify and classify a rotator cuff tear, based on the segmentation of the tear from MR images, was developed and validated. A neural network was trained to segment the rotator cuff tear from an MR image and automatic methods for calculating tear width and retraction and for classifying the tear according to pattern, extension and retraction were implemented. The accuracy of the automatic segmentation and the automated tear analysis were evaluated relative to the ground truth of manual segmentations by a clinical expert, and analyzed based on the ground truth segmentations. Variance in the manual segmentations was assessed in a interrater variability study of two clinical experts. The error of the automatic segmentation to one of the two clinical experts are meant to be in the same region. To make quantification accessible the whole pipeline was implemented in an existing webapp. The results were also evaluated clinically by intraoperative measurements of the rotator cuff tear performed on a separate dataset of six patients. The accuracy of the tear retraction calculation based on the developed automatic tear segmentation was 6.56 mm ± 6.48 mm in comparison to the interrater variability of tear retraction calculation based on manual segmentations of 3.77 mm ± 3.58 mm. These results show that an automatic quantification of a rotator cuff tear is possible. The achieved accuracies of the quantification pipeline for rotator cuff tears need to be improved to make them clinically useable. The large interrater variability of manual segmentation based measurements, highlights the difficulty of the tear segmentations task in general. To improve the accuracy of an automated segmentation, a larger dataset for training may be required or a semi-automatic approach could be used.