YOLOv8 algorithm-aided detection of patellar instability or dislocation on knee joint MRI images.

Abstract

Background: Patellar instability (PI) or patellar dislocation (PD) is challenging to diagnose accurately based on medical history and clinical manifestations alone. While X-ray, computed tomography (CT), and magnetic resonance imaging (MRI) are commonly employed for detecting PI or PD, computer vision has not yet been widely utilized for this purpose.

Purpose: To explore the feasibility of computer vision, specifically the You Only Look Once (YOLO) algorithm, in identifying patellar instability or dislocation.

Material and methods: A total of 550 patients (190 diagnosed with patellar instability or dislocation) were divided into a training set (n = 360), validation set (n = 90), and external test set (n = 100). Four indicators were measured on transverse knee MRI scans to determine the presence of patellar instability, and 450 images were labeled using Labelme software. YOLO version 8 (YOLOv8) was refined using these labeled images and validated on 100 unlabeled images. The diagnostic accuracy of YOLOv8 was compared with that of a junior radiologist.

Results: The sensitivity, specificity, and accuracy of the refined YOLO model and the junior radiologist were 62%, 97%, and 83%, and 62%, 82%, and 74%, respectively. Although the YOLO model demonstrated slightly higher accuracy, the difference did not reach statistical significance (P = 0.093). The YOLO model required approximately 14.01 ± 10.34 ms to interpret each image, significantly shorter than the 9.55 ± 2.39 s required by the radiologist (P < 0.001).

Conclusion: The refined YOLOv8 model is not inferior to junior radiologists in identifying patellar instability or dislocation and offers a significantly faster interpretation time.