{"id":26999,"date":"2026-09-27T20:02:20","date_gmt":"2026-09-27T18:02:20","guid":{"rendered":"https:\/\/www.itspy.cz\/thesis\/deep-learning-methods-for-histological-image-processing\/"},"modified":"2026-10-23T13:46:05","modified_gmt":"2026-10-23T11:46:05","slug":"deep-learning-methods-for-histological-image-processing","status":"publish","type":"thesis","link":"https:\/\/www.itspy.cz\/en\/thesis\/deep-learning-methods-for-histological-image-processing\/","title":{"rendered":"Deep learning methods for histological image processing"},"content":{"rendered":"<p>Histopathological whole-slide images (WSIs) are hard to<br \/>\nanalyze and computationally demanding. Diagnosis relies on<br \/>\nrelationships between local and global context, and<br \/>\nclinical use requires interpretability. The research<br \/>\nhypothesis asks whether hierarchical transformers with<br \/>\nexplicit relational attention can improve multi-scale WSI<br \/>\nmodeling and explainability. We propose the Relation-Aware<br \/>\nHierarchical Image Pyramid Transformer (RA-HIPT), which<br \/>\nfuses whole-slide and patch embeddings via cross-attention<br \/>\nwith graph regularization and a custom loss function aimed<br \/>\nat relation discovery. We performed evaluations under self-<br \/>\nand weak supervision on controlled synthetic WSIs and<br \/>\nBRACS. For evaluation, we used attention maps,<br \/>\nregion-to-region graphs, class-relation visualizations, and<br \/>\nconfusion analysis. A synthetic pipeline combines VQ-VAE,<br \/>\nsemantically conditioned diffusion, and inpainting-based<br \/>\nstitching. Empirically, VQ-VAE reached an FID of 13.14,<br \/>\nwhile conditioned sampling on BCSS with DDIM-50 achieved an<br \/>\nFID of 122.84 and a mean SSIM of 0.13. Alongside this, we<br \/>\nperformed seam ablations and a qualitative assessment of<br \/>\nstructures with 15 participants. With RA-HIPT, heatmaps and<br \/>\nrelation graphs highlighted tissue structures and their<br \/>\nconnections, which were validated by the pathologist.<br \/>\nDownstream classification tasks showed that RA-HIPT<br \/>\nachieved, on average, a ~16% increase in accuracy over<br \/>\nfrozen HIPT. These findings demonstrate synthetic dataset<br \/>\nquality, model interpretability, and the model\u2019s ability to<br \/>\ncapture biologically relevant structures.<\/p>\n","protected":false},"featured_media":26761,"template":"","meta":{"_acf_changed":false,"_links_to":"","_links_to_target":""},"university":[212,210],"thesis-year":[466],"class_list":["post-26999","thesis","type-thesis","status-publish","has-post-thumbnail","hentry","thesis-year-466"],"acf":{"autor":"Patrik Kozl\u00edk","portret":"","vedouci":"prof. Ing. Vanda Bene\u0161ov\u00e1, PhD."},"_links":{"self":[{"href":"https:\/\/www.itspy.cz\/en\/wp-json\/wp\/v2\/thesis\/26999","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.itspy.cz\/en\/wp-json\/wp\/v2\/thesis"}],"about":[{"href":"https:\/\/www.itspy.cz\/en\/wp-json\/wp\/v2\/types\/thesis"}],"version-history":[{"count":1,"href":"https:\/\/www.itspy.cz\/en\/wp-json\/wp\/v2\/thesis\/26999\/revisions"}],"predecessor-version":[{"id":27054,"href":"https:\/\/www.itspy.cz\/en\/wp-json\/wp\/v2\/thesis\/26999\/revisions\/27054"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.itspy.cz\/en\/wp-json\/wp\/v2\/media\/26761"}],"wp:attachment":[{"href":"https:\/\/www.itspy.cz\/en\/wp-json\/wp\/v2\/media?parent=26999"}],"wp:term":[{"taxonomy":"university","embeddable":true,"href":"https:\/\/www.itspy.cz\/en\/wp-json\/wp\/v2\/university?post=26999"},{"taxonomy":"thesis-year","embeddable":true,"href":"https:\/\/www.itspy.cz\/en\/wp-json\/wp\/v2\/thesis-year?post=26999"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}