A two-stage Convolutional neural network for meniscus segmentation and tear classification

Artificial intelligence for interpreting MRI meniscal tear could prioritize high risk patients and assist clinicians in making diagnoses. In this study, a two-stage end-to-end convolutional neural network, with Mask rcnn as backbone for object detecting and Resnet for classification, is proposed for automatically detecting torn in the meniscus on MRI exams. With training dataset of 507 MR images and validation dataset of 69 MR images, the meniscus detection achieves a recall of 0.95 when 1 false positive of 1 image, and the ROC for classification of torn meniscus get a AUC of 0.99.