Multi-task learning for automated recess detection and distension classification in hemophilic patients

Abstract

Haemophilic patients can experience joint bleeding that if not promptly treated can lead to hemophilic arthropathy. Hence early detection of bleeding is essential to prevent permanent damage to the articulations. The use of ultrasound imaging has become increasingly popular in diagnosing joint recess distention resulting from joint bleeding. Nonetheless, there is currently a lack of computer-aided diagnostic tools available to assist practitioners in the diagnosis of such conditions. Our research addresses this issue by developing an automated system that can detect the subquadricipital recess of the knee and determine whether it is distended using ultrasound images. We propose two approaches to detect and assess the distension of the subquadricipital recess: the first approach uses a one-stage object detection algorithm, while the second approach employs a multi-task approach. Our experimental evaluation with 483 annotated images shows that the object detection-based approach achieves a balanced accuracy of 0.74 with a mean IoU of 0.66, while the multi-task approach achieves a higher balanced accuracy of 0.78 with a slightly lower mean IoU value. In addition, we conducted an evaluation to determine the model’s ability to identify distension when caused by hemarthrosis. The model achieves a sensitivity of 77% on an excerpt of 27 images.