STOLÁRIK, S. Hluboké neuronové sítě pro detekci landmarků v obraze [online]. Brno: Vysoké učení technické v Brně. Fakulta informačních technologií. 2024.
Mr. Stolárik approached his Bachelor’s Thesis diligently and with enthusiasm. He displayed a clear interest in both deep learning and its application in the field of medical imaging. He exhibited a great ability to comprehend the methodologies outlined in scientific articles, identify their limitations, establish a training environment for various experiments, conduct them, and assess the results reasonably. Working with Samuel Stolárik on his final project has been very rewarding, and I look forward to continuing our collaboration to refine and optimize the last proposed method.
| Kritérium | Známka | Body | Slovní hodnocení |
|---|---|---|---|
| Informace k zadání | In his work, Mr. Stolárik focused on the still relevant issue of automatically detecting landmarks in cephalometric X-ray images. The significance of this task is underscored by the fact that the foremost medical image computing conference has held a challenge on this topic for the past two years. The complexity of the work lies in examining various representations of the input data and training labels sourced from scientific papers, often including somewhat unconventional approaches, such as utilizing graph neural networks for image analysis. This involved preparing several datasets with various data and label representations and conducting experiments with methods based on various approaches. The assignment was completed without any concerns. | ||
| Práce s literaturou | The student carefully searched and studied numerous high-quality scientific articles published in recent years. In addition, he explored many articles that weren't directly necessary for the successful fulfilment of the defined frame of the solution, yet they helped him grasp the bigger picture of the overall topic. When he was given additional literature recommendations by the supervisor throughout the year, he managed to comprehend the suggested works with impressive ease. | ||
| Aktivita během řešení, konzultace, komunikace | Mr. Stolárik regularly and frequently attended consultations on the progress of his work during both semesters. He came prepared with specific questions about the methodologies presented in the research papers. His curiosity clearly indicated that he was very interested in the subject. | ||
| Aktivita při dokončování | The work was completed on time, allowing me to provide feedback on the final version. My remarks are reflected in the submitted version of the technical report. | ||
| Publikační činnost, ocenění | Not known. |
The student became well-versed in the problem of detecting cephalometric landmarks using deep learning. He proposed his own modifications of convolutional networks and demonstrated their benefits. Its results are not too far from the state-of-the-art methods from the MICCAI Automatic Cephalometric Landmark Detection Challenge in 2023, a great result of Batchelor's thesis. I have some comments about the technical report, but overall, I consider the work excellent.
| Kritérium | Známka | Body | Slovní hodnocení |
|---|---|---|---|
| Náročnost zadání | In the case of deep learning and neural networks, the bachelor's student does not have much connection to the taught courses and has to acquire the necessary knowledge and good practices on his own. | ||
| Rozsah splnění požadavků zadání | All assignment points are met. Beyond the scope of the assignment, the student experimented with both classical convolutional neural networks and a graph neural network, which is an extension of the original assignment. | ||
| Rozsah technické zprávy | Most chapters have an ideal scope, and everything is discussed sufficiently and without unnecessary details. Only Chapter 2.2.1, Landmark Detection in Echocardiographs, is too detailed because the author does not address this task. Instead, I would appreciate more details and drawings of the architecture of models that were successful in the MICCAI challenge. | ||
| Prezentační úroveň technické zprávy | 80 | The technical report is well-structured, straightforward and easy for the reader to understand. A few minor comments: It's pretty awkward to state in Table 4.1 that the metric used is described later; I should know what the numbers mean when I look at the table. Some sections of the proposed solution and the implementation fit more into theoretical chapters and descriptions of existing methods, e.g. 4.3.1 Regression of Isotropic Gaussian Heatmaps or 4.4.2 Graph Neural Networks. The list of specific values for random perturbation (p. 30) and used hyperparameters (Table 5.1) should have been in the experiments, not the implementation. | |
| Formální úprava technické zprávy | 85 | The report is written in clear English without mistakes and typos. The typographical and linguistic aspects of the work are very good. I found perhaps only one unreferenced image. Figures 2.4 and 2.6 are raster and poor quality, and Figure 4.7 is unnecessarily large. Also, I do not recommend changing the ratio of width and height for medical images, which distorts anatomical structures (Figure 6.10 or the poster in the appendix). | |
| Práce s literaturou | 95 | Study literature is chosen appropriately and touches on the topic of the work. These are primarily scientific articles from which the author visibly drew inspiration for his methods and experiments. | |
| Realizační výstup | 95 | In addition to the results of the experiments, the implementation output consists of Python scripts that work with data and train various models. The adopted code part of the EchoGlad model is marked correctly. I appreciate that the student conceived the practical implementation as a small framework for training networks and thought about its structure. The codes are clear and commented on. | |
| Využitelnost výsledků | The student was inspired by existing methods and proposed his own modification of the U-Net model for detecting cephalometric landmarks called CHugNet, where he uses the regression of heatmaps, which do not have an isotropic character. He also experimented with random perturbation techniques in network training. His experiments show that both methods are beneficial. |
eVSKP id 156814