USTSINAU, U. Segmentace nádorů mozku v MRI datech s využitím hloubkového učení [online]. Brno: Vysoké učení technické v Brně. Fakulta elektrotechniky a komunikačních technologií. 2020.
The student proposed a deep learning approach for segmentation of tumors in MRI images. The List of Contents and sorting of the Chapters is consistent and well designed. A formal and technical sound of the thesis is on a good level. The student referenced relevant literature and used it appropriately. The student used consultations and came up with specific questions to the extent necessary, thus demonstrating his ability to work independently. The student proposed and tested three different algorithms and he carried out the analysis of its training quality based on a similarity of images in the training dataset. Achieved results are interesting, but I am missing some objective comparison with some methods published by other authors. Due to this drawback, one task from the assignment of the thesis is not completely solved. Considering these facts, I evaluate the work with a grade D (60 points).
The submitted diploma thesis deals with the topic of segmentation of brain tumors in MRI images. In the first part of the thesis, the author describes the basic theory of tomographic imaging systems and the possibilities of their use for imaging in neurology. The author also describes the theory of convolutional neural networks, which are further used for processing of MRI data. In the theoretical part, I lack a broader literature review of methods used directly on the assigned topic. In the experimental part, the student designed and implemented a method of segmentation of brain tumors and tested it on a publicly available BRATS database. Experiments connected with training as well as testing of algorithms are described in detail. The results of segmentation are clearly discussed. However, as part of the graphic presentation of the results, I would appreciate a wider discussion what are the main benefits and drawbacks of the proposed methods. It would also be useful to show a comparison of segmentation results together with ground-truth data. Further, I miss a quantitative comparison of segmentation results with other authors. Given that BRATS database was created as part of a conference challenge, I believe that the results of other authors should not be difficult to find. Moreover, comparison of achieved results with other authors is included in the thesis assignment - point 5). As this point of the assignment is missing in the thesis, I recommend to include the discussion of comparison of achieved results with other authors in the final presentation during the thesis defense. However, despite the above-mentioned issues and the missing point of the assignment, I believe that the student has achieved satisfactory results, as evidenced by his work. It is clear that the student had to study the extensive issues of convolutional neural networks and especially the possibility of implementation in the processing of a large image datasets. Rating: D/60 points.
eVSKP id 126757