Towards Graph-based Diffusion Models in Digital Pathology
Event details
| Date | 20.06.2023 |
| Hour | 10:00 › 12:00 |
| Speaker | Manuel Monteiro Lança Madeira |
| Location | |
| Category | Conferences - Seminars |
EDIC candidacy exam
Exam president: Prof. Pierre Vandergheynst
Thesis advisor: Prof. Pascal Frossard
Thesis co-advisor: Dr. Dorina Thanou
Co-examiner: Prof. Maria Brbic
Abstract
Graph deep learning approaches to digital pathology data have been particularly successful due to their capability of capturing complex dependencies between tissue entities, such as cells. The usage of deep generative models in this setting holds great promise by enabling the augmentation of the scarce digital pathology data and providing a more interpretable framework for the analysis of the reproduced biological mechanisms.
Motivated by these premises, we first analyse a paper that leverages Graph Neural Networks to capture and characterize disease-relevant microenvironments, showcasing the advantages of such techniques to digital pathology. We proceed to the introduction of diffusion models through the analysis of the paper that laid its empirical foundations as a state-of-the-art approach in generative modelling. Then, we consider the adaptation of diffusion methods to discrete state-spaces, setting a first step towards the unification of the two former papers, i.e., graph- based diffusion models. Finally, we briefly discuss open research directions that promise to further improve the generation of interpretable and biologically plausible data, as the incorporation of biological priors and hierarchical generation schemes.
Background papers
Exam president: Prof. Pierre Vandergheynst
Thesis advisor: Prof. Pascal Frossard
Thesis co-advisor: Dr. Dorina Thanou
Co-examiner: Prof. Maria Brbic
Abstract
Graph deep learning approaches to digital pathology data have been particularly successful due to their capability of capturing complex dependencies between tissue entities, such as cells. The usage of deep generative models in this setting holds great promise by enabling the augmentation of the scarce digital pathology data and providing a more interpretable framework for the analysis of the reproduced biological mechanisms.
Motivated by these premises, we first analyse a paper that leverages Graph Neural Networks to capture and characterize disease-relevant microenvironments, showcasing the advantages of such techniques to digital pathology. We proceed to the introduction of diffusion models through the analysis of the paper that laid its empirical foundations as a state-of-the-art approach in generative modelling. Then, we consider the adaptation of diffusion methods to discrete state-spaces, setting a first step towards the unification of the two former papers, i.e., graph- based diffusion models. Finally, we briefly discuss open research directions that promise to further improve the generation of interpretable and biologically plausible data, as the incorporation of biological priors and hierarchical generation schemes.
Background papers
- Denoising Diffusion Probabilistic Models (https://proceedings.neurips.cc/paper_files/paper/2020/file/4c5bcfec8584af0d967f1ab10179ca4b-Paper.pdf)
- Structured Denoising Diffusion Models in Discrete State-Spaces (https://proceedings.neurips.cc/paper_files/paper/2021/file/958c530554f78bcd8e97125b70e6973d-Paper.pdf)
- Graph deep learning for the characterization of tumour microenvironments from spatial protein profiles in tissue specimens (https://www.nature.com/articles/s41551-022-00951-w)
Practical information
- General public
- Free