Moderní inteligentní metody segmentace periosteálního svalku z klinických RTG obrazů

Abstract

This diploma thesis deals with the automatic segmentation of periosteal callus in wrist X-ray images using deep learning. The theoretical part summarizes bone healing, the periosteum, periosteal reaction, and the basic principles of medical image segmentation. The practical part focuses on data preparation from the GRAZPEDWRI-DX dataset, conversion of annotations into binary masks, and the design of segmentation models based on the U-Net architecture. A total of 2,235 images with corresponding masks were used in the experiments. Several network configurations and different image preprocessing methods were tested. Among the models trained on raw data, the best results were achieved by model v3 with an average Dice coefficient of 0.6857. The overall best performance was achieved by model PREPROC_v1 using local histogram equalization, with an average Dice coefficient of 0.6982 and IoU of 0.5561. The results showed that suitable architecture settings and image preprocessing can improve segmentation quality. The thesis confirms that automatic segmentation of periosteal callus is technically feasible, although it remains challenging, especially in small and low-contrast lesions.