A Few-shot approach to MRI-based Knee Disorder Diagnosis using Fuzzy Layers

Mohana Singh, K. Ram Prabhakar, Viswanath Pamulakanty Sudarshan, Murali Poduval, Arpan Pal, Jayavardhana Gubbi

Abstract: Magnetic Resonance Imaging (MRI) has a long-standing acceptance as the de facto technology for obtaining high-resolution three-dimensional images of soft tissue pathologies of the knee. However, the volume and amount of information in MRI sequences warrant a need for automated interpretations that can assist radiologists in making faster and more consistent diagnoses. Despite the popularity of deep learning in achieving this goal, it is limited by the dependency on large amounts of data. This work explores the detection of knee injury classes from a data-efficiency perspective. It also attempts to incorporate soft decision-making that is prevalent in the medical diagnosis domain. First, a training regime utilizing knowledge transferred from related diagnoses and a pool of unlabeled data is proposed. Secondly, to aid the deep-learning-based model with soft computing elements, novel fuzzy layers are employed along with the proposed training regime. Experiments are performed on multiple base models and compared with existing baselines. Results show that the proposed approach elevates the base models’ performance substantially compared to all the baselines.