Learning Fourier descriptors for computer-aided diagnosis of the supraspinatus

Supraspinatus muscle disorders are frequent and debilitating, resulting in pain and a limited range of shoulder motion. The gold standard for diagnosis involves an invasive surgical procedure. As part of a proposed clinical workflow for noninvasive computer-aided diagnosis (CAD) of the condition of the supraspinatus, we present a method to classify three-dimensional shapes of the muscle into relevant pathology groups, based on magnetic resonance (MR) images.

We obtained MR images of the shoulder from 72 patients, separated into five pathology groups. The imaging protocol ensures that the supraspinatus is consistently oriented relative to the MR imaging plane for each scan. Next, we compute the Fourier coefficients of two-dimensional contours lying on parallel imaging planes and integrate the corresponding frequency components across all contours. To classify the shapes, we learn the Fourier coefficients that best distinguish the different classes.

We show that our method leads to significant improvement when compared to previous work. We are able to distinguish between normal shapes and shapes that possess a pathology with an accuracy of almost 100%. Moreover, we can differentiate between the different pathology groups with an average accuracy of 86%.

We confirm that analyzing the three-dimensional shape of the muscle has potential as a form of diagnosis reinforcement to assess the condition of the supraspinatus. Moreover, our proposed descriptor based on Fourier coefficients is able to distinguish the different pathology groups with accuracies higher than those obtained by previous work, indicating its potential application to support a system for CAD of the supraspinatus.