[Data-Driven Biomedical Science Team Seminar] Talk by Prof. Thomas Walter on "Predictive Models in Computational Pathology"

Title: Predictive Models in Computational Pathology

Speaker: Prof. Thomas Walter (Mines Paris, PSL University / the Centre for Computational Biology, PSL University / Institut Curie)

Abstract:
Whole slide images (WSI) are microscopy images of stained tissue sections. They are very large and complex, often displaying millions of individual cells, their environments, and the overall tissue structure. WSI are routinely used in cancer treatment centers for diagnosis, patient stratification, and treatment selection.
Computational pathology is the field concerned with the automatic analysis of WSI The main workhorse of Computational Pathology is Deep Learning (DL). However, due to the size and complexity of WSI, standard methods in DL, such as Convolutional Neural Networks cannot be used directly: more advanced techniques such as Multiple Instance Learning in conjunction with Self-Supervised Learning need to be applied. Scarcity of annotated datasets and potential biases are further pitfalls in the use of DL in this context.
Here, I will give an overview of our recent work in the field of Computational Pathology. First, I will present our recently published approaches for automatic grading of head and neck and uterine cervix cancer. Grading is an important task in pathology, impacting patient care and treatment choice. As we show, computer aided grading can help to overcome inter-grader variability. Second, I will show applications where we aim at predicting molecular features from WSI, in particular the prediction of Homologous Recombination Deficiency (HRD), a defect in the DNA repair pathway that sensitizes cancer cells for treatment with PARP inhibitors. Prediction of HRD from WSI could provide an interesting way of preselecting patients with high probability of benefiting from this treatment. Finally, I will present our latest development in the field: Giga-SSL, a strategy we have recently developed to (1) improve performance of DL models on small datasets (e.g. rare cancer subtypes) and (2) ease WSI analysis by providing powerful representations that can easily be integrated into bioinformatics workflows. We provide such WSI encodings for the entire TCGA as a resource that can be used to confront molecular and morphological analyses.

References:
1. Lubrano, M.; Bellahsen‐Harrar, Y.; Berlemont, S.; Atallah, S.; Vaz, E.; Walter, T.; Badoual, C. Diagnosis with Confidence: Deep Learning for Reliable Classification of Laryngeal Dysplasia. Histopathology 2023, his.15067. https://doi.org/10.1111/his.15067
2. Lubrano, M.; Lazard, T.; Balezo, G.; Bellahsen-Harrar, Y.; Badoual, C.; Berlemont, S.; Walter, T. Automatic Grading of Cervical Biopsies by Combining Full and Self-Supervision. In Computer Vision – ECCV 2022 Workshops; Karlinsky, L., Michaeli, T., Nishino, K., Eds.; Springer Nature Switzerland: Cham, 2023; pp 408–423. https://doi.org/10.1007/978-3-031-25082-8_27.
3. Lazard, T.; Bataillon, G.; Naylor, P.; Popova, T.; Bidard, F.-C.; Stoppa-Lyonnet, D.; Stern, M.-H.; Decencière, E.; Walter, T.; Vincent-Salomon, A. Deep Learning Identifies Morphological Patterns of Homologous Recombination Deficiency in Luminal Breast Cancers from Whole Slide Images. Cell Reports Medicine 2022.
4. Lazard, T.; Lerousseau, M.; Decencière, E.; Walter, T. Giga-SSL: Self-Supervised Learning for Gigapixel Images. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (CVPR) workshops; 2023; pp 4304–4313.
5. Lazard, T.; Lerousseau, M.; Gardrat, S.; Vincent-Salomon, A.; Stern, M.-H.; Rodrigues, M.; Decencière, E.; Walter, T. Democratizing Computational Pathology: Optimized Whole Slide Image Representations for The Cancer Genome Atlas; preprint (biorxiv), 2023.

Bio:
Thomas Walter is full professor at Mines Paris and director of the Center for Computational Biology, a research group at Mines Paris, and part of the joined unit between Institut Curie, Mines Paris and INSERM on Computational Oncology. His research focuses on the development and application of methods in Machine Learning and Computer Vision to biological and medical image data.
Thomas Walter received his engineering degree from Saarland University in 1999 and completed his PhD thesis at the Centre for Mathematical Morphology at Mines Paris in the field of Computer Aided Diagnosis of Diabetic Retinopathy, a severe and frequent eye disease. He then joined the Ellenberg lab at the European Molecular Biology Laboratory (EMBL), where he developed Computer Vision methods to analyze the first genome-wide RNAi screen by live cell imaging in a human cell line in order to identify genes required for cell division. In 2012, he joined the Centre for Computational Biology at Mines Paris, first as an independent researcher and since 2018 as director. His work focuses on Bioimage Informatics with applications to High Content Screening and Computational Pathology. Since 2019, he holds a chair at the PaRis Artificial Intelligence Research InstitutE (PRAIRIE), and he is a full professor of Mines Paris since 2021. He is also active in teaching Deep Learning for Bioimage Analysis in a number of courses. He has published more than 80 articles (h-index: 32).