Přístupnostní navigace
E-application
Search Search Close
Master's Thesis
Author of thesis: Ing. Tibor Kubík
Acad. year: 2022/2023
Supervisor: doc. Ing. Michal Španěl, Ph.D.
Reviewer: prof. Dr. Ing. Pavel Zemčík, dr. h. c.
Deep neural network architectures designed for traditional signals like regularly sampled images and grids do not straightforwardly translate to irregularly sampled and triangulated surfaces, point clouds, and other geometric representations. As the acquisition tools that produce such 3D data are becoming broadly available, this generalization is increasingly needed for treatment planning in digital medicine. This work aims to examine the application of deep learning techniques for the analysis of triangular meshes. A recurrent multi-view approach is proposed for the task of segmentation of teeth in surface dental scans, an industry-desirable automation. On complex real-world orthodontic cases containing dental irregularities or scanned appliances, the proposed method outperforms both the conventional segmentation algorithm based on 3D Graph-Cut, and non-Euclidean methods that analyze point clouds or directly meshes. It achieves an average weighted IoU score of 0.966 and Hausdorff distance at 95 percentile of 0.382 mm. The results are promising for a deployment in dental planning software, enabling clinicians to streamline their workflow and devote more attention and focus on the treatment itself.
mesh segmentation, teeth segmentation, dental scans, geometric deep learning, recurrent multi-view neural networks, ConvLSTM, PointNet, PointNet++, MeshCNN
Date of defence
24.08.2023
Result of the defence
Defended (thesis was successfully defended)
Grading
A
Process of defence
Student nejprve prezentoval výsledky, kterých dosáhl v rámci své práce. Komise se poté seznámila s hodnocením vedoucího a posudkem oponenta práce. Student následně odpověděl na otázky oponenta a na další otázky přítomných. Komise se na základě posudku oponenta, hodnocení vedoucího, přednesené prezentace a odpovědí studenta na položené otázky rozhodla práci hodnotit stupněm A.
Topics for thesis defence
Language of thesis
English
Faculty
Fakulta informačních technologií
Department
Department of Computer Graphics and Multimedia
Study programme
Information Technology and Artificial Intelligence (MITAI)
Specialization
Computer Vision (NVIZ)
Composition of Committee
prof. Ing. Adam Herout, Ph.D. (předseda) doc. Ing. Martin Čadík, Ph.D. (člen) Ing. František Grézl, Ph.D. (člen) Ing. Michal Hradiš, Ph.D. (člen) Ing. David Bařina, Ph.D. (člen) doc. Mgr. Adam Rogalewicz, Ph.D. (člen)
Supervisor’s reportdoc. Ing. Michal Španěl, Ph.D.
Tibor Kubík put a lot of energy and enthusiasm into his diploma thesis what is evident on its exceptional extent and quality. He became a great expert in the field. His publishable research results also practically usable in clinical practice prove that he is also very capable of understanding real clinical problems and applying his theoretical knowledge in practice. He proved his very good technical skills and attention to detail in the implementation and experiments. I really enjoyed supervising this thesis!
Thematically, the assignment follows bachelor's thesis where Tibor Kubík focused on detection of landmark points on dental scans, and it is also a continuation of FIT's partnership with the TESCAN 3DIM company for which the segmentation of dental scans more robust against various atypicall cases is a very attractive topic.
Successfully solving the assignment required studying the latest trends in deep learning for 3D mesh processing, adaptation of published state of the art methods to the given task and a large amount of experimental work.
The extent to which the student fulfilled the assignment exceeds normal expectations for a diploma thesis. In addition to experiments with so-called multi-view approach, he expanded the work with experiments with most recent architectures like PointNets and MeshCNNs what required a lot of changes because of their very different nature.
The student searched for and studied a number of recently published scientific papers and he became an expert on the current state of knowledge in the topic of processing 3D models using various types of neural networks like common convolutional ones and graph neural networks.
The student devoted himself to the topic with great interest and regularly consulted his proposals for next steps. A large amount of work was completed in the winter semester before his departure for an Erasmus stay abroad. Nevertheless, he still found time to further expand his experiments with other architectures.
The work was completed on time and my minor comments on carefully prepared technical report were addressed.
We plan to publish the results of the work in a scientific journal.
Grade proposed by supervisor: A
Reviewer’s reportprof. Dr. Ing. Pavel Zemčík, dr. h. c.
The presented work is very well done. The text of the work is nicely structured and readable, written in a good English and the experimental results of the work based on the software created by the student are very good as well. Overall, the work is simply excellent.
Evaluation level: značně obtížné zadání
The assignement of the presented diploma thesis has benn. to my opinion, quite difficult as it required extended self-study by the student as well as extensice experimental work.
Evaluation level: zadání splněno a práce obsahuje podstatná rozšíření
To my opinion, the assignement of the thesis has not only been fulfilled but it was exceeded as the student experimentally evaluated more possible approaches to the solution.
Evaluation level: přesahuje obvyklé rozmezí
The text of the thesis is unusually long and it contains in total 74 pages (plus annexes).
The presentation level of the thesis text is very good, its structure is logical and it is well understandable for readers. Perhaps a bit more attention could have been paid to the titles of the subchapters some of which are not that explanatory when looking at the Contents of the text.
On the formal side, the work is done excellently, except for minor imperfections, such as e.g. quite empty pages (e.g. 2, 30...) some of which with "text orphan" (e.g. 2).
The work with literature is excellent, the work contains 80 references, most of which in pefect format.
The algorithm and software created as a result of the work is excellent and to my opinion fully functional.
The results are original and they are, to my opinion, very well exploitable in healthcare applications for teeth segmentation.
Grade proposed by reviewer: A
Responsibility: Mgr. et Mgr. Hana Odstrčilová