Predicting Discus Hernia from MRI Images Using Deep Transfer Learning

The capacity to timely detect and classify discus hernia in individuals means faster access to adequate therapy. Standard way to diagnose the patients is through magnetic resonance images (MRI), which uses axial and sagittal view. Previous research has revealed that transfer learning is a useful approach when it comes to small datasets. In this paper we investigate the use of deep learning models to identify the level (healthy, L4/L5, L5/S1) and the side (healthy, bulging, left, right, center) of discus hernia in patients from MRI images. Dataset used consisted of combined publicly accessible and restricted local database of 1169 MRI images in sagittal view and 557 images in axial view. A board-certified radiologist manually classified images which was used as a golden standard. Several well-known convolutional neural networks were used in combination with transfer learning (i.e. VGG16, VGG19, DenseNet121, Xception). The results reveal competitive accuracy, as well as other metrics such as sensitivity, specificity, precision etc. Although the acquired performance is quite positive, additional investigations on larger datasets are necessary to get more robust conclusions.