GAVOROVÁ, Z. Traktografie axonálních svazků založená na multi-tensorových modelech [online]. Brno: Vysoké učení technické v Brně. Fakulta elektrotechniky a komunikačních technologií. 2017.
Bachelor Zuzana Piskořová has regularly and actively consulted the given task during the whole period with me and other national and international consultants, and has demonstrated the ability to work independently. Within the literature review, she proved that she is capable of working with a large volume of the recent literature available mostly in English only. Although she is presenting the negative result and although the theoretical part is much larger than the practical part, she has fulfilled all the required points of the assignment. Only for the 6th point, I have small complaints listed below. The inconsistency between the theoretical and practical volume is given by the fact, that she had to learn, understand and apply the large volume of theoretical knowledges out of the scope and level of our master degree programme, where some of them were derived within the last year. We have entered the task to her to solve the filtered tractogram optimizing problem with the non-square matrix pseudo-inversion as the method converging to the global optimum. She is suggesting several ideas why the task is not solvable, especially on classic personal computer. And here, some my complaints arise, while I find some evaluations for insufficient. 1. The task did not have to be solved on PC. The student is erudite enough to be able to use national grid computational center, get sufficient computational requirements and compute the pseudo-inversion for the matrix of orders 10^10 to validate precisely the computational time and roundoff error. 2. I am missing the filtered tractogram visualizations for both optimizing methods (COMMIT and pseudo-inversion). Because the original tractogram can be considered for the gold standard, it could be qualitatively and quantitatively assessed that COMMIT works and pseudo-inversion fails for matrices of the high order. 3. The main purpose of the tractogram filtering is to filter out the white matter tracts with non-homogeneous microstructural properties (e.g. axonal density or axonal diameter), i.e. the intra-axonal compartment. But the used multi-compartment model ball-zeppelin-stick does not allow to model any mentioned microstructural variability over different tracts. So, I find the chosen model as innapropriate. Formally, the thesis is written and divided well. It contains only several tweaks in equations which slightly decrease the value of the presented work (e.g eq. 5.3 vector x is not defined; or page 46 variable I_p, the p is not limited). The most crucial issue is the bad written and not-updated abstract. Overall, the level of the thesis belongs to category very hard. And, the submitted diploma thesis is at the level of the working progress doctoral thesis, i.e. in the middle-time, which could be successfully solved in following time sufficient for its solution. For that reasons, I recommend the thesis for the defense and evaluate it 80/B.
The master thesis deals with diffusion MRI (dMRI), namely modeling and quantification of diffusion and the task of tractography. The student has performed a fairly extensive and good literature review of basic dMRI principles, available single-voxel diffusion models and tractography approaches. A review of available software libraries, as required by the instructions (point 2), was however very short and insufficient (page 36). The student chose to implement an alternative optimization algorithm of the existing COMMIT tractography method. The student has proposed to use pseudo-inversion based on singular value decomposition. It is not clear from the text to what extent the complete COMMIT algorithm has been implemented and what software packages have been used. It seems that a testing dMRI dataset has been preprocessed only to estimate the size of the matrix to be pseudo-inversed. Then the proposed pseudo-inversion method (implemeted using an available library for python) has been applied only on small artificial matrices to infer the theoretical computational demands for the testing dMRI dataset. The conclusion was that the chosen algorithm is not suitable for the given problem due to the ill-posed nature of the problem and memory demands. This conclusion was sufficiently well substantiated. The level of English is good in the theoretical part, although with often incorrect use of articles. In the experimental part, the level of English was substantially worse. With respect to the complex character of the topic I recommend the thesis for the defense.
eVSKP id 102375