Detection and Classification of Knee Ligament Pathology based on Convolutional Neural Networks

Nowadays, medical imaging has almost completely replaced traditional techniques for diagnosis and treatment planning due to their non-invasiveness, speed and ease of manipulation. Due to their popularity the medical imaging pushed the bottleneck towards the medical staff that, post-acquisition, have to analyse each case. This process is not only slow but also subjective to a level that it becomes error prone. The anterior cruciate ligament (ACL) is one of the most injured ligaments of the knee. Injuries occur predominantly in a young and sports-active population. Many patients are left with significant disability following injury to the ACL. The injury leads to alteration in the mechanics of the knee. Because some of the cases require surgery it is imperative to be able to classify and detect ACL pathology with high accuracy. Hence this paper studies the usage of pre-trained convolutional neural networks with residual connections (ResNet), in conjunction with image processing techniques to detect ACL pathology and distinguish between multiple tear levels. The ResNet used is not the vanilla ResNet but rather a modified version adapted to work with 3D volumes composed of 2D images (slices). With an obtained accuracy of 87% it is safe to say that the model's output can be successfully used by radiologists to set a 1st baseline diagnostic without any effort almost instantly.