Combining Deep Learning Techniques and Image Analysis Methods for Vertebrae, Disc, and Spinal Stenosis Characterization in Lumbar Spine MRI

Artificial neural networks (ANN) in medicine are presented as a decision-making support method that allows obtaining more optimal solutions in terms of time and resource management. In this article, an ANN-aided decision support method has been developed to classify and quantify low back disease, specifically vertebrae and disc characterization, and spinal stenosis from MRI. A set of 1960 slices from 200 patients extracted from T2-weighted lumbar spine MRI has been used to train and evaluate neural networks. The segmentations of the discs and vertebrae have been accomplished using U-nets. Experiments on T2-weighted MR images of 200 subjects show that U-nets achieve performances with mean Dice similarity coefficients of 0.79 and 0.76 for the segmentations of 10 vertebrae and 9 intervertebral discs, respectively.Clinical Relevance-A computer-aided diagnosis (CAD) methodology using U-Net for segmentation of vertebrae and intervertebral discs on lumbar spine MRI with minimal user input is proposed. The suggested approach holds great potential as a clinical support system for radiologists.