A Machine Learning Approach for the Objective Sonographic Assessment of Patellar Tendinopathy in Collegiate Basketball Athletes

Patellar tendinopathy (PT) is a knee injury resulting in pain localized to the patellar tendon. One main factor that causes PT is repetitive overloading of the tendon. Because of this mechanism, PT is commonly seen in "jumping sports" like basketball. This injury can severely impact a player's performance, and in order for a timely return to preinjury activity levels early diagnosis and treatment is important. The standard for the diagnosis of PT is a clinical examination, including a patient history and a physical assessment. Because PT has similar symptoms to injuries of other knee structures like the bursae, fat pad, and patellofemoral joint, imaging is regularly performed to aid in determining the correct diagnosis. One common imaging modality for the patellar tendon is gray-scale ultrasonography (GS-US). However, the accurate detection of PT in GS-US images is grader dependent and requires a high level of expertise. Machine learning (ML) models, which can accurately and objectively perform image classification tasks, could be used as a reliable automated tool to aid clinicians in assessing PT in GS-US images. ML models, like support vector machines (SVMs) and convolutional neural networks (CNNs), use features learned from labelled images, to predict the class of an unlabelled image. SVMs work by creating an optimal hyperplane between classes of labelled data points, and then classifies an unlabelled datapoint depending on which side of the hyperplane it falls. CNNs work by learning the set of features and recognizing what pattern of features describes each class. The objective of this study was to develop a SVM model and a CNN model to classify GS-US images of the patellar tendon as either normal or diseased (PT present), with an accuracy around 83%, the accuracy that experienced clinicians achieved when diagnosing PT in GS-US images that were already clinically diagnosed as either diseased or normal. We will also compare different test designs for each model to determine which achieved the highest accuracy.

GS-US images of the patellar tendon were obtained from male and female Virginia Tech collegiate basketball athletes. Each image was labelled by an experienced clinician as either diseased or normal. These images were split into training and testing sets. The SVM and the CNN models were created using Python. For the SVM model, features were extracted from the training set using speeded up robust features (SURF). These features were then used to train the SVM model by calculating the optimal weights for the hyperplane. For the CNN model, the features were learned by layers within the CNN as were the optimal weights for classification. Both of these models were then used to predict the class of the images within the testing set, and the accuracy, sensitivity and precision of the models were calculated. For each model we looked at different test designs. The balanced designs had the same amount of diseased and normal images. The designs with Long images had only images taken in the longitudinal orientation, unlike Long+Trans, which had both longitudinal and transverse images. The designs with Full images contained the patellar tendon and surrounding tissue, whereas the ROI images removed the surrounding tissue.

The best designs for the SVM model were the Unbalanced Long designs for both the Full and ROI images. Both designs had an accuracy of 77.5%. The best design for the CNN model was the Balanced Long+Trans Full design, with an accuracy of 80.3%. Both of the models had more difficulty classifying normal images than diseased images. This may be because the diseased images had a well defined feature pattern, while the normal images did not. Overall, the CNN features and classifier achieved a higher accuracy than the SURF features and SVM classifier. The CNN model is only slightly below 83%, the accuracy of an experienced clinician. These are promising results, and as the data set size increases and the models are fine tuned, the accuracy of the model will only continue to increase.