Diagnosing Knee Injuries from MRI with Transformer Based Deep Learning

Magnetic Resonance Images (MRI) examinations are widely used for diagnosing injuries in the knee. Automatic interpretable detection of meniscus, Anterior Cruciate Ligament (ACL) tears, and general abnormalities from knee MRI is an essential task for automating the clinical diagnosis of knee MRI. This paper proposes a combination of convolution neural network and sequential network deep learning models for detecting general anomalies, ACL tears, and meniscal tears on knee MRI. We combine information from multiple MRI views with transformer blocks for final diagnosis. Also, we did an ablation study which is training with only CNN, and saw the impact of the transformer blocks on the learning. On average, we achieve a performance of 0.905 AUC for three injury cases on MRNet data.