Combined Image Enhancement, Feature Extraction, and Classification Protocol to Improve Detection and Diagnosis of Rotator-cuff Tears on MR Imaging

抄録

Purpose: The diagnosis of most rotator cuff tears (RCTs) relies upon magnetic resonance (MR) imaging, but direct capture of MR images without enhanced image processing leads to poor image contrast and potential misdiagnosis. Therefore, we developed a 2-stage model for the detection and diagnosis of injury of the supraspinatus tendon.
Methods: The first stage used coupled weighted histogram separation (WHS) to improve image enhancement, and the second stage extracted suspicious texture, features of both spatial and spectral domains, and sequential floating forward selection (SFFS) selected features conducive to classification of RCTs. We then tested injuries of the supraspinatus tendon using the classifier.
Results: The extraction of features by SFFS can increase detection of supraspinatus injury by reducing the input vector by 57.78% from the enhanced input images. The receiver operating characteristic (ROC) curve indicated an azimuth (Az) value of 84.38% when SFFS selected 76 features to construct a support vector machine (SVM) classifier from the enhanced images, compared with 56.94% when all 180 features from the raw input images were used for the construction.
Conclusions: The performance of the classifier constructed by SFFS-selected features is superior to that using all features. These findings can serve as references to improve diagnosis and treatment of supraspinatus injuries.