Student se ve své práci zabýval velmi aktuálním tématem, a to sice 3D zobrazováním pomocí difrakční tomografie. Oproti standardnímu přístupu (kdy se využívá koherentní zdroj světla) se student zabýval možnostmi difrakční tomografie pomocí nízko-koherentního osvětlení. V rámci své práce si připravil vzorky (fantómy buněk, vytvořené dvoufotonovou polymerací), samostatně provedl experiment na vyvíjeném prototypu tomografického mikroskopu a sám napsal poměrně náročný program na 3D rekonstrukci obrazu. Z formálního hlediska je předložená práce napsaná velmi dobrou angličtinou, výklad je doprovázen pěknými obrázky. Co by se práci dalo vytknout je poněkud slabší diskuse nad dosaženými výsledky. Nicméně, práce jednoznačně splňuje všechny kritéria kladené na závěrečnou práci a já studentovi přeju všechno nejlepší v jeho následující kariéře.

 1. Standard requirements

The thesis „3D Imaging in a Coherence–Controlled Holographic Microscope” written by Bc. Jozef Buček under supervision of Mgr. Jakub Bělín, Ph.D., fulfils the standard requirements for Master’s thesis.

2. Objectives

The overarching goal of the thesis is to develop pilot software for 3D image reconstruction in optical diffraction tomography based on the coherence–controlled digital holographic microscope platform developed in FME BUT, as well as to validate it’s performance using simulated and experimental data. The methodology and results presented in the thesis align well with the objectives and support the claim, that the goal has been achieved.

3. Methodology

The applied methodology is quite extensive and appropriate for the task. First, the reader is introduced to the tomographic principles and its implementation via optical diffraction tomography, followed by hardware platform description, mathematical formalism of forward and inverse models, as well as the particular implementation of the reconstruction algorithm being the centrepiece of the thesis. The details on the hardware platform are sparse, however, it is well-justified by the fact that the main focus of the thesis is on the reconstruction algorithm, where the description of theoretical framework and implementation details are thorough and very well done. Since I am not an expert in implementation of tomographic solvers, I can’t comment on the accuracy of the mathematical tools presented in the thesis, however, the concepts of tomography under Born/Rytov scattering approximations, resulting Fourier Diffraction Theorem, as well as forward and inverse models (BPM, MLB, gradient descent method, loss function and regularization term) are presented and explained exceptionally well. They are also in agreement with the literature and widely used in this field. Validation of the developed algorithms is performed in simulations and experimentally, using custom imaging targets and cross-validated with another imaging instrument, which is essentially the gold standard in such works. Therefore, the overall methodology is well-designed and appropriately presented in the thesis.

Shortcoming of the methodology section is overreliance on the reference [32] (https://en.wikipedia.org/wiki/Optical_tomography) and reference [10] (HAISCH, Christoph. Optical Tomography. Annual Review of Analytical Chemistry. 2012, vol. 5, no. 1, pp. 57–77. issn 1936-1335.), which is also recommended in the “Further reading” section on the Wikipedia page. This Wikipedia article heavily influenced Section 1 of the thesis, but unfortunately is of poor quality and not so relevant to the ODT or QPI in general. The page itself also has a disclaimer indicating multiple issues with completeness and sources. This resulted in somewhat disjointed introduction of optical tomography in the context of this work and highlights the need for better source selection. Similar article (https://en.wikipedia.org/wiki/Tomography) seems like a much better place to start the search for primary and secondary sources, and citing Wikipedia directly should generally be avoided in academic writing.

4. Results

The Results section presents the work done between the two institutes, one located in Czechia and the other in Poland, which is advantageous and provides opportunity to utilize additional expertise. The section starts with the introduction of the phantom design based on the phase images of the cells and its fabrication. Then, the implementation details of the two tomographic forward and inverse solvers are described. It is not clear whether they were implemented from scratch or based on available, open-source implementation, which is a shortcoming. Simulated reconstructions are described in detail, however, there is no discussion over hyperparameter selection, therefore their values are not that useful and influence on the reconstruction are undocumented. Quantified error metric (MSE) has been utilized to express the accuracy of the reconstruction, followed up with the discussion and qualitative analysis, which are convincing. The figures presenting the comparison between the initial and reconstructed 3D RI are unfortunately lacking, as they are all shown in different scales and direct comparison is not possible. The same dynamic range and a plot would go a long way to better present the results. The discussion is focused on direct comparison of MBP and MLB, while the discussion on the general performance related to the expected system parameters (e.g. theoretical spatial resolution vs imaged feature size) and quality comparison with similar reconstructions presented in the literature are also lacking. Similarly in the experimental section, the overall procedure is well-documented and described, but the results and discussion are lacking critical overview and conclusions on what could be improved. Some additional examples where implementation of the methodology could be improved include:

- Only reconstructions with multi-stage, iterative algorithms are shown. It would be very useful to see the basic reconstruction obtained with raw data using simple inversion under FDT to establish the baseline. As previously mentioned, it is not clear whether reconstruction errors in simulations/experiments arise from erroneous input data, developed algorithms or hyperparameter selection.

- (page 36) “The elongation along the z-axis is significantly greater in the reconstructions produced by our algorithm. This is mainly due to the smaller number of projections available for our algorithm compared to the GTVIC reconstruction algorithm. The number of available projections has a direct impact on elongation: the fewer projections available, the greater the elongation along the optical axis.” – this is generally incorrect, as more projections beyond ~20 primarily improve signal-to-noise and have no ability to fill the missing cone region in the K-space resulting in the elongation. While the iterative methods use this redundant data to better fill the missing region, the most efficient solution in my opinion would be to use both datasets to reconstruct using both algorithms to isolate variables.

- (page 36 and fig. 4.9) It is not clear what is the issue with the phase wrapping and how range convention (0;2pi or -pi;pi) leads to shown issues. Phase can be unwrapped and extended beyond this range using dedicated algorithms.

- Discussion of the fabrication errors, especially voxel size and its influence on the size of the cavity.

Finally, the conclusion section summarizes the results by mostly reiterating the content of each section. This would be a perfect place to discuss the developed algorithms and results in the broader context of state of the art, highlight most important results and successful aspects of the work, and most importantly what could be done in hindsight to improve the methodology. Discussion of possible further work is also warranted, but unfortunately also missing. These are all relatively small shortcomings, but taken together they add up and make the results feel incomplete.

5. Literature, typesetting, technical writing and visuals

The thesis is written well from linguistic and technical standpoint, with clear sections, numbered equations and figures, as well as references. The content is grouped logically and relatively well, but the separation between the background knowledge, specific instruments and techniques utilized in the thesis, and work done specifically by the student should be more clear and thought out. Key differences, system parameters, reconstruction times, MSE values etc. should be gathered in tables to highlight them and facilitate comparison. Figures are prepared with various level of care, e.g. fig. 1.2 does not convey any useful information, panels of fig. 3.1 are in completely different colormaps with no annotations, and fig. 4.5, 4.6, 4.7 4.10, and 4.12 are also shown in different scales without section plots or comparison with expected distributions, limiting their utility. The bibliography is extensive and well-utilized, recent with most work from 2010-2020. Minor issues with references include citing Wikipedia and introducing optical tomography from narrow, unexpected perspective as mentioned earlier, as well as references were not numbered in the order of their appearance in the text. Topics for thesis defence:

1. What do you think is the biggest source of error in your BPM and MLB algorithms, and why?
2. Have you tried using your algorithms on publicly available QPI datasets?
3. With your current knowledge and experience, which parts of the methodology would you consider well done, and which would you do differently, and how?