Využití neuronových sítí pro detekci a hodnocení endometriálních lézí s využitím interpretovatelných metod umělé inteligence

Abstract

This master’s thesis deals with the use of convolutional neural networks for the automatic segmentation of pelvic abnormalities and endometrial lesions from magnetic resonance imaging scans. The aim of the thesis was to analyze the available dataset of pelvic MRI images, design a suitable segmentation method, and complement it with an interpretation of the model’s decision-making using explainable artificial intelligence methods.

Selected approaches to intensity normalization were compared in the thesis. Based on the preprocessed data, two segmentation models using the nnU-Net architecture were trained, namely for the segmentation of pelvic abnormalities and for the segmentation of endometrial lesions. The models were evaluated using the DSC, IoU, HD95, and ASSD metrics.

For the segmentation of pelvic abnormalities, the final model achieved a mean DSC of 0.557, IoU of 0.459, HD95 of 14.4 mm, and ASSD of 2.9 mm. For the segmentation of endometrial lesions, the achieved results were lower. The mean DSC reached 0.412 for fold 0, 0.000 for fold 1, and 0.461 for fold 2. The best results were achieved in cases with larger and more regularly shaped structures, whereas the segmentation of small or atypical lesions was generally less successful.

The Grad-CAM method was used to interpret the model’s decision-making. In the most successfully segmented cases, the mean activation value inside the reference region was 0.282, whereas outside the reference region it was 0.022. In less successful segmentations, these values were 0.171 and 0.036. Thus, not only a quantitative evaluation of the segmentation was achieved, but also a complementary interpretation of the model’s decision-making.

The results confirm that the proposed approach is applicable to the segmentation of pathological structures in MRI images; however, the accuracy of endometrial lesion segmentation is strongly affected by the limited size of the dataset and by the variability in lesion appearance.