Multi-class Classification of Ultrasonic Supraspinatus Images based on Radial Basis Function Neural Network

This article reports a study on applying texture analysis methods for classifying the different rotator cuff disease groups which are normal, tendon inflammation, calcific tendonitis and rotator cuff tear by using ultrasonic images. In conventional diagnosis, physicians observe the micro/macro structures of ultrasonic tendon images to judge the severity of rotator cuff disease. The accuracy of visual observation depends on the expertise of physicians. It is often not reliable. The supraspinatus is usually involved in the above-mentioned diseases progression categories. Four texture analysis methods, gray-level co-occurrence matrix, texture spectrum, fractal dimension and texture feature coding method, were used to extract features of tissue characteristic of supraspinatus. In the feature selection stage, two different criteria, the mutual information selection and the F-score measurement were independently used to select powerful features and then compare them. It revealed that features selected by the two different methods were not significantly different, and could potentially be reliable for classification. Meanwhile, radial basis function networks were also designed to discriminate test images into one of the four disease groups in the classification stage. The percentage of correct classification was more than 92.5% using this proposed automatic computer system. Experimental results show that the proposed method performed very well for the classification of ultrasonic supraspinatus images.