The analysis of orthodontic anomalies in 3D dental scans is a challenging task with clear practical relevance. An automatic classification model could assist clinicians in better understanding planned cases and could also provide useful structured information for other algorithms during 3D digital planning. The difficulty of the task comes from two main aspects: the non-trivial application of deep learning methods to 3D shapes, and the common challenges of medical datasets, such as limited data availability and class imbalance.

The goal of the thesis was to experimentally evaluate several general-purpose frameworks for 3D shape classification and compare their performance under different settings. An additional goal was to explore the explainability of the predictions, so that clinicians could obtain visual feedback indicating which parts of the 3D model contributed to the model’s decision. Mr. Dedič fulfilled the assignment, and I am satisfied with the achieved results.

Not known.

The student consulted the work several times. The meetings usually focused on presenting the completed parts of the work and validating the next steps. The student was prepared for the meetings.

A higher frequency of consultations might have led to even more interesting results, as the student clearly has strong potential. Nevertheless, the collaboration was smooth, and the agreed direction of the work was followed.

Evaluation level: zadání splněno

All required points of the assignment were met exhaustively. The student compared three state-of-the-art models for classifying 3D point clouds, performed an extensive ablation study to optimise their tuning and interpret the properties of various variants from the perspective of the selected classification tasks, and attempted to visually interpret the results using three XAI methods.

The technical report is well written, and it is clear what the author experimented with and how. As a reader, I would appreciate a few changes:

• Chapter 5 mixes specific implementation details and crucial experimental design. I would find it more understandable to formulate experimental questions (Chapter 5.7) right at the beginning in a separate chapter together with the overall overviews of ​​the solution (Figures 5.1-5.4).
• The number of experiments is impressive; however, the sections devoted to the ablation study are difficult to read, making it hard to keep track of crucial findings. It would be appropriate to summarise only the most interesting findings in the report (i.e. the resulting hyperparameter settings and pipeline setup) and put detailed tables and comments in the appendix. It is unnecessary to list numbers in tables and repeat them in the text.

The study literature is broad, covering the necessary knowledge and state-of-the-art approaches. The author draws heavily on scientific papers.

The work extends the results of other authors and provides additional insights into the accuracy of dental attribute classification using state-of-the-art point cloud models. An interesting aspect of the work is the use of various XAI techniques to interpret model behaviour, an area that remains underexplored in 3D shape analysis. It will be possible to build on the results in the future.