Research Article: Quantitative diagnosis of rotator cuff tears based on sonographic pattern recognition

Date Published: February 28, 2019

Publisher: Public Library of Science

Author(s): Ruey-Feng Chang, Chung-Chien Lee, Chung-Ming Lo, Terry K. Koo.


The lifetime prevalence of shoulder pain is nearly 70% and is mostly attributable to subacromial disorders. A rotator cuff tear is the most severe form of subacromial disorders, and most occur in the supraspinatus. For clinical examination, shoulder ultrasound is recommended to detect supraspinatus tears. In this study, a computer-aided tear classification (CTC) system was developed to identify supraspinatus tears in ultrasound examinations and reduce inter-operator variability. The observed cases included 89 ultrasound images of supraspinatus tendinopathy and 102 of supraspinatus tear from 136 patients. For each case, intensity and texture features were extracted from the entire lesion and combined in a binary logistic regression classifier for lesion classification. The proposed CTC system achieved an accuracy rate of 92% (176/191) and an area under receiver operating characteristic curve (Az) of 0.9694. Based on its diagnostic performance, the CTC system has promise for clinical use.

Partial Text

Rotator cuff disorders are the most common cause (up to 70%) of shoulder pain [1], with a lifetime prevalence approaching 70% [2]. The financial burden of shoulder pain on the United States health care system is estimated at $7 billion annually [3], and the substantial loss of productivity is often underestimated. The mechanisms of rotator cuff diseases are believed to possess a dynamic pathology, with subacromial impingement as the initial stage and rotator cuff tear as the final stage [4]. Rotator cuff disorders include tendinopathy, calcific tendinitis, tears, bursitis, and bursal reactions [5]. Among these disorders, rotator cuff tears, which have a prevalence rate of 20.7%, are the most severe forms [6]. Individuals suffering from rotator cuff tears may have severe shoulder pain, weak forward elevation, abduction or external rotation, which can detrimentally affect the activities of daily life.

Tables 1 and 2 show whether the proposed image features, including intensity and texture features, can be significant in tear classification. As a result, four intensity and 11 texture features obtained a statistically significant p-value less than 0.001. After feature selection, the relevant image features were selected and combined in the classifier to generate a prediction model. Three performance results of the CTC system based on different feature sets are shown in Table 3. After backward elimination, three of four intensity features including Mean, Skewness, Kurtosis were selected and combined in the classifier. The intensity feature set attained an accuracy of 91%, a sensitivity of 92%, and a specificity of 91%. For texture features, Correlation, Information measure of correlation, and Inverse difference normalized were selected to be the most relevant according to their combination performance. The texture feature set attained an accuracy of 89%, a sensitivity of 89% and, a specificity of 89%. Benefiting from complementary advantages, the combined intensity and texture feature sets including selected Mean, Kurtosis, Inverse difference normalized, and Inverse difference moment achieved an accuracy of 92%, which is better than using intensity and texture feature sets individually.

The proposed CTC system based on intensity and texture features was established to interpret tissue echogenicities of shoulder ultrasound images. The prediction model built by a logistic regression classifier achieved an accuracy of 92% for identifying rotator cuff tears and tendinopathies. The high accuracy suggests that the proposed CTC system is useful for assessing the presence of rotator cuff tears. The classification result was obtained via leave-one-out cross-validation due to the limited cases. The accuracy presented in this study provides us a direction that the proposed CTC system works well in tear classification while the morphology features are useless for differentiation in the observation. With respect to the selected features, Tears tend to be darker due to its higher value of mean intensity and centralized with higher kurtosis value. Besides, high Inverse difference normalized, and Inverse difference moment mean the gray-scale distribution is uniform and lacking variance.




Leave a Reply

Your email address will not be published.