Ultrasound-based knee osteoarthritis detection using instance segmentation and classification

Knee osteoarthritis (OA) is a degenerative joint disease that causes knee pain, chronic symptoms, limited mobility, and a reduced quality of life. Early detection of knee OA is crucial because timely intervention can help slow disease progression and improve treatment effectiveness. Ultrasonography provides a non-invasive imaging modality for assessing soft tissue abnormalities associated with knee OA, although the complexity of ultrasound images can make accurate analysis challenging. This study proposes a two-stage framework integrating instance segmentation and classification for ultrasound-based knee OA detection. In the first stage, instance segmentation was performed to identify soft tissue regions from ultrasound images acquired from the medial, lateral, and suprapatellar views. The segmentation results were then used to generate segmentation-based masked images that preserve relevant anatomical information for subsequent classification. In the second stage, the generated images were used to classify knee OA conditions. The proposed framework was implemented using the YOLOv11 architecture. The segmentation results demonstrated the ability of the framework to segment soft tissue structures across different ultrasound imaging views. In the medial view, the model obtained an IoU of 0.69 and an mAP of 0.74. In the lateral view, the model obtained an IoU of 0.65 and an mAP of 0.83. In the suprapatellar view, the model obtained an IoU of 0.81 and an mAP of 0.99. The classification results showed that using segmentation-based masked images yields better performance than using the original images. The model achieved an accuracy of 0.97, a precision of 0.93, a recall of 0.93, and a specificity of 0.98 on the segmented images. These findings indicated that integrating instance segmentation and classification improved the performance of ultrasound-based knee OA detection.