MR Image Based Self-Supervised Learning for Knee Injury Classification

The cost of getting quality images and the cost of expertise which view these images have created barriers in medical imaging applications. Often, supervised machine learning approaches require large annotated datasets to reach the desired level of performance. We present a self-supervised learning (SSL) approach that exploits spatial anatomical structures in magnetic resonance (MR) video frames for the diagnosis of knee-related problems to overcome these shortcomings. We use a pretext model that extracts fixed features that are consistent across spaces. The features we extract from the pretext model competes very well for the multi-label classification task in the class imbalance setting. Notably, the model representation of even minority classes is efficient without using the common dataset imbalance management techniques. This project report study offers an in-depth exploration of the methodologies and steps undertaken.Self-supervised learning has become a powerful method for efficiently learning representations, no longer relying on human-labelled data. Our approach uses a self-supervised framework for classifying distinct types of knee injuries in MR images using Knowledge Graphs. Along with tackling multi-label classification, we investigate the limitations of popular pretext tasks (such as predicting geometric transformations) in order to gain a better understanding of the scope and limitations of self-supervised learning.