Introduction: Artificial intelligence is increasingly transforming the field of medical diagnostics, especially in the processing of medical images. One of the key procedures in this area is segmentation, which enables the precise delineation of anatomical structures. With the development of deep learning methods, new possibilities for automatic, faster, and more reproducible segmentation have emerged. In this thesis, we examined the use of the 3D Slicer software and the MONAI Label extension for automatic segmentation of prostate magnetic resonance images and compared its performance with reference-based segmentation. Purpose: The master's thesis addresses the use of artificial intelligence for prostate magnetic resonance image segmentation in 3D Slicer with the MONAI Label tool and compares the results with manual and semi-automatic segmentation. The aim is to evaluate the efficiency and execution time of different approaches and to determine the advantages of artificial intelligence in the recognition of anatomical structures. The results are expected to demonstrate how artificial intelligence can accelerate workflows and enhance the clinical applicability of segmentation in radiology. Methods: In this thesis, we applied both descriptive and experimental methods. First, through a literature review, we analyzed existing approaches to prostate magnetic resonance image segmentation. In the experimental part, we performed manual, semi-automatic, and automatic segmentation of fifteen T2-weighted prostate magnetic resonance scans. The results were compared with reference segmentations previously performed by experienced radiologists. The performance of each method was evaluated using the Dice similarity coefficient and segmentation time, with a focus on the central gland and peripheral zone. Among other things, we also measured prostate volume and performed statistical analysis of the data. Results: The results revealed differences in accuracy and time efficiency among manual, semi-automatic, and automatic segmentation methods. Manual segmentation showed the highest similarity to the reference data but was also the most time-consuming. Automatic segmentation achieved high Dice coefficient values and minimal volumetric deviations, while being significantly faster than the other methods. The semi-automatic method proved to be the least accurate, with the largest deviations in volume estimation. Overall, the findings indicate that automatic segmentation provides the best balance between accuracy and speed, particularly in the segmentation of the central gland, while accuracy in the peripheral zone was slightly lower across all methods. Discussion and conclusion: The results of this study confirm that artificial intelligence plays a significant role in the segmentation of prostate magnetic resonance images. The automatic method using MONAI Label proved to be the fastest and most volumetrically accurate, achieving the highest Dice coefficient values. Manual segmentation was too time-consuming for routine use, while the semi-automatic method was the least reliable. We conclude that automatic segmentation enables efficient, reproducible, and user-friendly image processing, representing an important step toward greater clinical applicability of artificial intelligence.
|
