Universität zu Köln

Sancéré, Lucas ORCID: 0009-0000-3857-5641 (2026). Multi-model deep learning pipeline and graph neural networks for skin cancer cells and tissues image analysis. PhD thesis, Universität zu Köln.

PDF (PhD Thesis of Lucas Sancéré) Thesis_Lucas_Sancere.pdf - Published Version Bereitstellung unter der CC-Lizenz: Creative Commons Attribution Non-commercial Share Alike. Download (57MB)

Abstract

Deep learning became an increasing popular field of research in the last 25 years. It is only in 2012 with DeepBind predicting binding preferences of DNA and RNA binding proteins and in 2015 with U-net architecture segmenting neural structures from electron microscopy that we saw major applications of this emergent field to biology. Since then, academia faced an exponential increase in the application of deep learning methods to biology and medicine. MICCAI conference, The Medical Image Computing and Computer Assisted Intervention Society, one of the largest conference for machine learning applied to medicine, received 756 submitted papers in 2016 and 3,667 in 2025.\medskip Our research is situated within this broad historical movement. This work focus on computer vision model for the analysis of medical images. More specifically on the analysis of Whole Slide Images (WSIs) of cutaneous Squamous Cell Carcinoma (cSCC). Whole Slide Images are high dimension megapixel images of tissue for which nucleus of cells are visible. They are commonly acquire in hospital's pathology departments for diagnostic or to follow treatment progress. Cutaneous Squamous Cell Carcinoma is a common type of skin cancer, highly prevalent worldwide. Our work methodology and software's are applied to this cancer as an example use-case but are design for further use on other cancer types. We first developed Histo-Miner pipeline to segment and classify all cell nuclei from cSCC WSIs. Following this step, the pipeline is used to calculate relevant tissue features and then summarize the WSI as few key numbers for downstream analysis. We first applied our Histo-Miner software to predict therapy response of patients undergoing anti-PD1 immunotherapy through analyses of their WSI recordings. In addition to provide solid classification performance, Histo-Miner provided a list of key features responsible for therapy response and insights into the underlying biology. We then applied Histo-Miner on WSIs cSCC from 3 clinical centers to analyze the most predictive tiles in classification of progression status (disease progression or no progression). A transformer-based multiple instance learning model was first used to classify the WSIs. Then Integrated Gradient Method was used to identify the most relevant patches and Histo-Miner was applied on these for segmentation and classification of the cell nuclei. After classification, the pipeline was employed to calculate tissue and cell based features. Several cell based features showed significantly different distribution between the two groups, for instance non-progressors maintained homogeneous tumor patterns, while progressors were characterized by a higher degree of integration between tumor cells and neighboring cell populations. Finally, using classical machine learning model XGBoost on the features calculated by Histo-Miner on the most representative patches yielded to high classification accuracy. Lastly, we used cell graph representations and graph Transformers neural networks to improve on cell nuclei classification for the case of epithelial cells. We compared image-based and graph-based approaches on WSI cell classification, on 2 distinct scenarios. The first scenario is the classification of all epithelial cells from a singe WSI. The second scenario is the classification of epithelial cells from WSI patches of different patients. We revealed that graph Transformers with linear complexity are better performing than state of the art image-based methods on both cases. Building cell graph representations from WSI and performing classification from these graphs instead of the original image lead to improved classification performance and significantly faster training and evaluation.

Item Type:	Thesis (PhD thesis)
Creators:	Creators Email ORCID ORCID Put Code Sancéré, Lucas UNSPECIFIED https://orcid.org/0009-0000-3857-5641 UNSPECIFIED
Contributors:	Contribution Name Email Author Sancéré, Lucas UNSPECIFIED
URN:	urn:nbn:de:hbz:38-804152
Date:	2026
Language:	English
Faculty:	Faculty of Mathematics and Natural Sciences
Divisions:	Faculty of Mathematics and Natural Sciences > Department of Mathematics and Computer Science > Institute of Computer Science
Subjects:	Natural sciences and mathematics
Uncontrolled Keywords:	Keywords Language Deep Learning English Computer Vision English Histology English Transformers English Graph Neural Networks English Cells English Cancer English
Date of oral exam:	24 April 2026
Referee:	Name Academic Title Bozek, Katarzyna Prof. Dr. Möllenhoff, Kathrin Prof. Dr. Lässig, Michael Prof. Dr.
Refereed:	Yes
URI:	http://kups.ub.uni-koeln.de/id/eprint/80415