JÓKAY, D. Analýza panoramatických rentgenových snímků pomocí hlubokého učení [online]. Brno: Vysoké učení technické v Brně. Fakulta informačních technologií. 2025.
Mr. Jókay successfully applied known machine learning methods to a relevant problem in medical imaging. He demonstrated a very good level of engagement, and I am confident that with even more curiosity and initiative, he could deliver an excellent work.
| Kritérium | Známka | Body | Slovní hodnocení |
|---|---|---|---|
| Informace k zadání | The goal of the thesis was to apply state-of-the-art machine learning algorithms to one of the tasks in dental X-ray image analysis. One such actively researched task is abnormality detection, as highlighted by its inclusion in the DENTEX challenge at MICCAI 2023, the leading conference in medical imaging. The student chose to explore the use of YOLO algorithms for this purpose, experimenting with various model versions and sizes, and additionally examining the impact of various transfer learning scenarios. In the supervisor’s opinion, the assignment has been successfully fulfilled. | ||
| Práce s literaturou | While the student engaged with the core literature relevant to the topic, broader exploration of alternative approaches was limited, which slightly narrowed the research perspective. | ||
| Aktivita během řešení, konzultace, komunikace | Throughout both semesters, the student maintained a high level of activity, presenting well-prepared progress updates at regular and frequent intervals. | ||
| Aktivita při dokončování | The student ensured that the work was completed well before the deadline, allowing me to provide suggestions on the technical report. | ||
| Publikační činnost, ocenění | Not known. |
Dávid Jókay designed a series of experiments with clearly formulated questions and evaluated the proposed experiments, trying to interpret the results. His work demonstrates that he is capable of conducting meaningful experiments and producing valuable results. Some experiments did not yield the expected improvement, but he most likely learned a lot from them, and this is a very good prerequisite for his future work, perhaps of a scientific nature.
| Kritérium | Známka | Body | Slovní hodnocení |
|---|---|---|---|
| Náročnost zadání | |||
| Rozsah splnění požadavků zadání | In accordance with the assignment, the author focused on anomaly detection (such as dental caries and impacted teeth) in X-ray dental scans. He used the available DENTEX dataset and existing YOLO models for object detection and designed and implemented a series of experiments, including the fusion of outputs from two models trained for a slightly different task. | ||
| Rozsah technické zprávy | |||
| Prezentační úroveň technické zprávy | 85 | The technical report is well-structured and clear. Chapter 2 provides a concise summary of the history of YOLO and R-CNN object detection models, as well as key differences between the generations. This overview may not be so essential for the work, but it is nicely written. Of the other models, only DETR is mentioned. Experiments are extensive, so I appreciate a clear summary of each group of experiments. A few minor critical comments: The main goal of the thesis is described in a paragraph (on page 21) within the chapter about the dataset. This is easy to overlook if the reader is familiar with the dataset and doesn't read the chapter carefully... Details of the MixUp and Mosaic augmentation techniques are missing; one must guess what they do from the figures. The author could have merged some of the sub-tables into one to comprehensively compare different variants of model training (e.g. Tables 5.2 and 5.3). | |
| Formální úprava technické zprávy | 90 | The technical report is typographically excellent and is written in very good English. | |
| Práce s literaturou | 75 | The student focused primarily on the YOLO family of models, which he studied in detail and experimented with. If he had explored other types of models more thoroughly, it would have earned the bachelor's thesis an excellent rating. | |
| Realizační výstup | 85 | The practical output of the work lies in experiments with YOLO models. The technical solution itself is relatively simple and is represented by Jupyter notebooks for running and evaluating experiments. I appreciate the use of the W&B framework for tracking and visualisation of ML training pipelines. | |
| Využitelnost výsledků | The experiments are sound, have a clear formulation of research questions which make sense. Some experiments may have been designed and realised slightly better, but overall, the experiments are wide and bring some useful insights into the performance of various YOLO models for the task of detecting dental tooth anomalies in X-ray images. An ablation study to show what particular augmentations are useful is missing. It is unclear whether the vertical flip and, particularly, advanced augmentations like the Mosaic are truly beneficial. To compare the models, one should find the best hyperparameters for each model, as each may require a slightly different setup. |
eVSKP id 162144