2025 ISAKOS Biennial Congress ePoster
Enhancing Ramp Lesion Detection In Acl Injury Through Deep Learning Technology
Hyung Jun Park, MD, PhD, Prof. KOREA, REPUBLIC OF
Gi Jun Shin, MD, Ansan-si, Gyeonggi-do KOREA, REPUBLIC OF
Sungwon Ham, PhD, Ansan-si, Gyeonggi-do KOREA, REPUBLIC OF
Euddeum Shim, MD, PhD, Prof., Ansan-si, Gyeonggi-do KOREA, REPUBLIC OF
Dong-Hun Suh, MD, PhD, Prof., Ansan, Gyeonggi KOREA, REPUBLIC OF
Jae Gyoon Kim, MD, PhD, Prof., Ansan, Gyeonggi KOREA, REPUBLIC OF
Korea University Ansan Hospital, Korea University College of Medicine, Ansan-si, Gyeonggi-do, KOREA, REPUBLIC OF
FDA Status Not Applicable
Summary
Deep learning combined with clinical risk factors shows potential to enhance MRI-based detection of ramp lesions in ACL injuries, approaching the diagnostic accuracy of expert radiologists.
ePosters will be available shortly before Congress
Abstract
Introduction
Accurate diagnosis of intra-articular injuries is crucial for managing anterior cruciate ligament (ACL) injuries, presenting significant challenges for clinicians. Ramp lesions, longitudinal tears at the meniscocapsular junction of the medial meniscus, are common in ACL injuries and exacerbate their severity by increasing anterior tibial translation and rotational instability. Clinical studies highlight the benefits of timely ramp lesion repair in restoring knee biomechanics. Despite MRI being a preferred diagnostic tool, its accuracy is suboptimal, with sensitivities ranging from 48% to 84%. This variability necessitates the use of clinical and radiographic risk factors, such as sex, age, bone marrow edema, and concurrent lateral meniscus tear, for preoperative diagnosis. The advent of artificial intelligence (AI) and deep learning technologies has shown promise in medical diagnosis, achieving results comparable to professional clinicians. This study aims to determine if deep learning technology can enhance MRI's diagnostic accuracy for detecting ramp lesions in ACL injuries and whether integrating clinical risk factors with processed image data improves diagnostic accuracy.
Methods
A retrospective study included 222 cases of patients who underwent surgery for ACL injuries, with exclusions for inadequate preoperative MRI. Demographic, surgical, and radiographic risk factors were analyzed, and MRI examinations were conducted using a 3.0-Tesla machine. Ramp lesions on MRI were defined as high signal changes or separations at the meniscocapsular junction. A deep learning model, utilizing the EfficientNetB0 architecture and AdamW optimizer, was developed for binary classification of ramp lesions. The model's performance was validated with training, validation, and test datasets, and clinical risk factors were incorporated to enhance accuracy. The accuracy, sensitivity, and specificity of the deep learning model were compared to the interpretations of the hospital's radiology experts.
Results
The accuracy of the deep learning model using MRI for detecting ramp lesions was inferior to that of expert radiologists. Accuracy was 73.3% versus 83.1%, sensitivity was 71.7% versus 77.4%, and specificity was 74.2% versus 87.3%, respectively. Among various risk factors, bone marrow edema on the posteromedial tibia (odds ratio, 4.112, p = 0.007) and concurrent lateral meniscus tear (odds ratio, 3.586, p = 0.011) were identified as clinically significant. Incorporating clinical risk factors improved the detection rate to a level similar to that of experts, with accuracy at 82.4% versus 81.9%, sensitivity at 71.5% versus 76.5%, and specificity at 81.2% versus 87.0%, respectively.
Discussion
The study presents a novel approach to improving ramp lesion detection in ACL injuries through deep learning, emphasizing the integration of clinical risk factors for enhanced diagnostic accuracy.