Achilles Tendinopathy Detection With Machine Learning Using Ultrasound Exams With Varied Quality And Resolution

PURPOSE: To enable a machine learning (ML) method that can be trained and applied directly using complete sets of Achilles tendon ultrasound images typically collected in a patient visit for Achilles tendinopathy (AT) diagnosis without the need to manually select and clean specific images, and to determine the impacts of varied image quality and image resolution on its detection performance. METHODS: A previously presented ML-based AT detection method is enhanced with the ability to automatically identify images that are more likely to present the pathology from within a set of Achilles tendon ultrasound images. The enhanced method uses an attention-based multiple instance learning approach to give greater emphasis on images that exhibit features of AT. Subsequently, images of different resolutions were synthetically produced using gaussian filters to represent the image quality of a wide variety of ultrasound machines in the field. The different resolution image sets are used to assess performance impacts. RESULTS: The study has collected 90 ultrasound exams of patients with and without AT. The enhanced ML model demonstrated a performance improvement from 96.33% to 97.89% area under the receiver operating characteristic (ROC) curve. In addition to the performance improvement, the model demonstrated the capability of identifying views within a patient exam that were more likely to present the pathology. When applied to lower resolution images, a slight (~5%) degradation in performance was seen with a gaussian kernel size of 15 - a resolution similar to a low-resolution hand-held ultrasound scanner we have access to. Further reduction in resolution was tested, causing the performance to continue to deteriorate. CONCLUSIONS: The enhanced ML method is shown capable of detecting AT using a set of ultrasound images acquired from a patient visit without the need of manual selection and cleaning. The method is therefore better suited for real-time feedback in clinical settings than other ML methods. The resolution study confirms that higher resolution ultrasound images do provide better AT detection performance, while reasonable performance can be achieved with even a lower resolution handheld ultrasound probe available today. Supported by: NIH #3P20GM121342-06S1, CU-PRISMA Innovation Fund Grant award.