Metody segmentace a identifikace páteře z RTG obrazových dat

Abstract

This diploma thesis describes segmentation of human spine from X-ray image data. The theoretical part summarizes essential knowledge about spine anatomy, common pathologies, different approaches of spine diagnostics and research of possible segmentation algorithms for this task. The practical part consists of designing and implementation of chosen segmentation algorithms in MATLAB software followed by evaluating it’s segmentation outputs, the used algorithms were convolutional neural network U-net and the active contour models. To compare the segmentations, a manual labels of spine from X-ray images were created to serve as a grand truth which represents ideal segmentation. The evaluation of results was based on selected evaluation metrics, the most relevant of which are F1 score and Jaccard index – IoU. In terms of accuracy, the active contour model provided significantly better results (average metrics values – F1 = 0,914; IoU = 0,842), as it adopted the spine structure very reliably and had consistent results across the entire dataset. The results of the neural network were decent enough but highly variable (average metrics values – F1 = 0,783; IoU = 0,656), since there were some images that the neural network coulnd’t segment properly. However, in terms of computational time, segmentation using the U-net neural network was many times faster and fully automatic.