A Clinical Decision Support System Based on Machine Learning: Toward High Detection Performance for Ligament Injury

Computer Vision (CV) is gaining traction in various fields of medicine, with one of its most significant applications being the development of medical diagnostic tools. In the domain of trauma and orthopedics, anterior cruciate ligament (ACL) tears, particularly among athletes such as footballers and skiers, are a common clinical problem. Most clinicians in this field have access to Magnetic Resonance Imaging (MRI), which is routinely performed by radiologists with specialized training in the interpretation of high‐definition knee scans. However, the diagnostic process is influenced by many factors that can compromise the reliability of clinical assessment, including cognitive fatigue, excessive time, and poor imaging modalities that reduce the reliability of the obtained images. This work proposes a diagnostic system to automate the analysis of MR image scans of the knee joints using Convolutional Neural Networks (CNNs). The system is designed to help practitioners automatically locate and analyze the ACL region and provide an assessment of pathological changes in the ligament. The solution is implemented with two connected CNNs, one dedicated to localizing the ACL within volumetric MR image data, and the other responsible for performing a classification that separates intact ligaments from those that are ruptured. The proposed methodology is validated on a curated dataset of 917 knee MR examinations obtained from the Rijeka Clinical Hospital Centre in Croatia. Each examination contains 32 sagittal slices with a spatial resolution of 320 by 320 pixels. A standard sequence of preprocessing steps was carried out to normalize the data prior to training in order to organize the input data. The entire classification pipeline is integrated into a diagnostic platform accessible via the web and Android, which aids clinical users in assessing the integrity of ACL injury in real time. Regarding the experimental results, the proposed system was able to achieve a classification accuracy of roughly 97.80%, well above the results achieved by traditional machine learning classifiers such as Support Vector Machines (SVM) and Random Forests (RF). The demonstrated efficacy of the system significantly strengthens its potential as a comprehensive tool to assist clinicians in evaluating ACL injuries and radiological evaluations of ACL injuries.