Transfer Learning and Data Augmentation in the Diagnosis of Knee MRI

John Haddadian, Mehala Balamurali

Abstract: Emergence of convolutional neural networks (CNNs) have offered better predictive performance and the possibility to replace traditional workflows with single network architecture. Recently developed MRNet CNN for the Knee MRI dataset has used AlexNet for their transfer learning implementation. This paper explores the effect of structural variations, data augmentation and various transfer learning implementations on the performance of a deep neural network in the classification task of knee MRI. Modifications of MRNet were generated by freezing the layers of the AlexNet backbone, replacing the backbone network AlexNet to other and applying the valid data augmentation techniques used on the dataset prior to input to the network. AlexNet based CNNs with layer-freezing achieved AUC for Abnormal, ACL lesion and Meniscal lesion classification of 0.913, 0.859, 0.792, an improvement over no layer freezing which had AUCs of 0.896, 0.842 and 0.773. although the result is less than that reported by Stanford’s AlexNet based classifiers of 0.937, 0.965 and 0.847 AUC. ResNet18 based classifier achieved AUCs of 0.843, 0.774, 0.671. VGG16 based classifier achieved AUCs of 0.728 0.690 0.711. Using color jitter for data augmentation resulted 0.938 AUC in abnormal classification.