Segmentace medicínských obrazů na základě učení bez učitele: laboratorní úloha

Abstract

This bachelor's thesis focused on regional segmentation of cluster analysis methods, where a total of three segmentation methods were tested on medical data from various imaging modalities. The data came from ultrasound imaging (US), X-ray imaging (RTG), computed tomography (CT) and, lastly, magnetic resonance imaging (MRI). The data were obtained from the Multiscan s.r.o. radio logy center in Pardubice and duly provided with a data sharing agreement. The data were completely anonymized in DICOM format, thus meeting the conditions for the protection and security of these patient data. The goal was to evaluate the performance and robustness of segmentation algorithms using cluster analysis, namely the K-means, Fuzzy C-means and K-medoids algorithms, where specific results were achieved using input parameters such as the number of clusters, distance metric and the selected number of iterations, which were compared with each other and, in the end, it was determined which method is suitable for segmenting certain medical data with implemented synthetic noise. The degradation of individual input data was performed using Salt and Pepper noise, Gaussian and Speckle noise. The evaluation was performed usingevaluation parameters that determined the influence of a certain type of noise of different intensity and also the number of clusters on the segmentation. The segmentation results were plotted in discrete graphs. The segmentation procedure was performed in the MATLAB software environment.