A Machine Learning Model for Automation of Ligament Injury Detection Process

Abstract

Good exploitation of medical data is very useful for patient assessment. It requires a diversity of skills and expertise since it concerns a large number of issues. Traumatic pathology is by far the most frequent problem among young athletes. Sport injuries represent a large part of these accidents, and those of the knee are the most important, dominated by meniscal and ligamentous lesions including that of the anterior cruciate ligament ($\mathcal {ACL}$). Magnetic Resonance Imaging ($\mathcal {MRI}$) is the reference for knee exploration, the number of knee MRI exams is in a perpetual increase thus of its contribution in the patient assessment and $\mathcal {MRI}$ machines availability. Therefore, radiologist’s time has become a limiting factor because of the large number of images to examine, in addition to the possibility of error in the interpretation. The possibility of automating certain interpretation functions is currently possible in order to limit the amount of errors and inter-observer variability. Deep learning is useful for disease detection in clinical radiology because it maximizes the diagnostic performance and reduces subjectivity and errors due to distraction, the complexity of the case, the misapplication of rules, or lack of knowledge. The purpose of this work is to generate a model that can extract $\mathcal {ACL}$ from $\mathcal {MRI}$ input data and classify its different lesions. We developed two convolutional neural networks ($\mathcal {CNN}$) for a dual-purpose, the first is to isolate the $\mathcal {ACL}$ and the second to classify it according to the presence or absence of lesions. We investigate the possibility of automating the $\mathcal {ACL}$ tears diagnostic process by analyzing the data provided by cross sections of patient $\mathcal {MRI}$ images. The analysis and experiments based on real $\mathcal {MRI}$ data show that our approach substantially outperforms the existing deep learning models such as support vector machine and Random Forest Model, in terms of injury detection accuracy. Our model achieved an accuracy rate equal to 97.76%.