Feasibility of Spine Segmentation in ML-Based Recognition of Vertebrae in X-Ray Images

Medicine remains one of the data-scarce domains, with typical datasets size not exceeding hundreds of thousands items. Correspondingly, straightforward application of end-to-end Deep Learning (ML) methods for Medical Imaging is often problematic. Carefully constructed pipelines that involve filter-based enhancement of raw images allow better outcome and leave more room for feasible data augmentation. In our paper, we explore and develop Computer Vision approaches for vertebrae recognition on a limited dataset of X-ray images of human spine. In particular, we test several pre-processing filters, of which Adaptive Histogram Equalization demonstrates the highest accuracy of 85.8% in a naïve model using the popular U-Net architecture. Further, we put forward the mechanism for the automated spine area cropping based on brightness cut-off. This additional step in the pipeline allowed improving the accuracy to 89.9%, which is comparable with the accuracy obtained for MRI images of higher quality. We plan to further increase the prediction quality of the model via spine segmentation into about 20 individual vertebrae and 8-fold data augmentation. The ultimate goal is improving the identification of the vertebrae corners, which currently has the relative error of 4.36%. Our intermediate results reported in the current paper might be of interest to ML engineers who pick the steps for X-ray spine images processing pipelines.