Addressing Class Imbalance for Transformer Based Knee MRI Classification

For assessing knee injuries, Magnetic Resonance Image (MRI) examinations are commonly utilized. Developing an automatic interpretable detection mechanism is an essential task for automating the clinical diagnosis of knee MRI. The imbalanced dataset problem is generally an issue for learning models in which the distribution of classes in the dataset is asymmetrical. The MRI datasets are generally imbalanced in favor of categories with injuries because patients who have an MRI are more likely to suffer a knee injury. Hence, it can be a challenging task to train a machine learning algorithm that can automatically handle class imbalance. In this paper, we propose both a network architecture and a comparison of the handling imbalanced dataset techniques to detect the general abnormalities in knee MR images. A network architecture that consists of CNN and transformer-based layers is proposed. Six different configuration methods for imbalanced data training are developed and compared with evaluation metrics (ROCAUC score, specificity, sensitivity, accuracy). Augmentation of additional data to the under-represented class and use of focal loss yield better classification specificity and AUC.