The TIA seminar is a regular event, organised by the Tissue Image Analytics Centre, that usually takes place on the 1st and 3rd Mondays of each month between 2pm and 3pm. We invite researchers and leaders in the area of computational pathology and associated areas to discuss their work and stimulate thought-provoking discussions.
Please contact the TIA Seminar Organiser Dr Simon Graham via email (Simon dot Graham at warwick dot ac dot uk) if you are interested in attending any of the seminars below.
|11th October 2021||
Memorial Sloan Kettering Cancer Center
|Mathematical Oncology Initiative|
|18th October 2021||
Pushpak Pati & Guillaume Jaume
IBM Research Zurich
|HistoCartography: Entity-graph representations and models in Computational Pathology|
|1st November 2021||
University Hospital of Zurich
|Deep Learning and the Arbitrary Orientation of Histology Images|
|15th November 2021||
Harvard Medical School
|6th December 2021||
Narmin Ghaffari Laleh
RWTH University Hostpital
|20th December 2021||
Çiğdem Gündüz Demir
Radboud University Medical Center
Memorial Sloan Kettering Cancer Center, NYC, USA (click here for a short bio)
11 Oct 2021
Title: Mathematical Oncology Initiative
Abstract: In this talk, I will present some of our recent work on addressing challenges in analysing H&E, IHC, and multiplex images. Specifically, I will talk about (1) advanced mathematical approaches (multi-marginal optimal mass transport) for stain normalisation and augmentation (as one way to achieve stain invariance), (2) advanced mathematical algorithms (vectorial optimal mass transport) for robust uni-/cross-modal (H&E/IHC/multiplex) WSI registration, (3) stain-invariant multitask deep learning framework, DeepLIIF, for simultaneous stain deconvolution/separation, cell segmentation, and quantitative single-cell IHC scoring (this framework is now deployed in Paige.ai's Slide Viewer), (4) interactive deep learning framework, ImPartial, for whole-cell segmentation and auto-thresholding in multiplex images, and (5) advanced mathematical (spectral analysis + optimal mass transport + Ricci curvature + non-abelian Fourier analysis) Spatial Profiling Toolbox for spatial biomarker quantification and discovery in digital pathology images. I will conclude with future directions and brief overview of our recent works in radiation oncology and surgery.
IBM Research, Zurich, Switzerland (click here for short bios)
18 Oct 2021
Title: HistoCartography: Entity-graph representations and models in Computational Pathology
Abstract: Computational pathology aims to provide computer-aided diagnosis tools to automate, assist and augment pathologists. To this end, we introduce you to HistoCartography that leverages biological entities and Graph Neural Networks for the modelling and analysis of histology images. In the first part, we will present a novel hierarchical cell-to-tissue-graph (HACT) representation for comprehensively encoding tissue regions. We will further introduce HACT-Net, a neural network to map HACT to histopathological breast cancer subtypes. The power of HACT-Net will be demonstrated on BRACS, a novel dataset for BReAst Cancer Subtyping of histology RoIs. In the second part, we will emphasize on the need for explainability to assist medical decisions and build trust between pathologists and AI algorithms. We will see how entity-graph processing can enable the explainability in pathologist-friendly terms. Finally, we will emphasize the scalability aspect of entity-graphs by introducing SegGini. It aims to semantically segment gleason patterns in prostate TMAs and WSIs by leveraging weak image-level supervision.
University Hospital of Zurich, Switzerland (click here for a short bio)
1 Nov 2021
Title: Deep Learning and the Arbitrary Orientation of Histology Images
Abstract: As deep learning has become the methodology of choice for image classification tasks, the robustness of deep learning models against irrelevant factors of variation is a necessary property for their safe deployment, especially in clinical environments. Solutions based on geometric deep learning have enabled models to gain invariance properties, thus bringing robustness guarantees against specific irrelevant variations. This talk will focus on the problem of the robustness to the arbitrary orientation of histology images in the context of supervised and unsupervised learning frameworks for computational pathology. Further, I will discuss the limitations of standard convolutional neural networks, how to expand existing architectures with rotation-equivariant operations and to what extent this is a solution.
Dr Faisal Mahmood
Harvard Medical School, Boston, USA
15 Nov 2021
RWTH University Hospital, Aachen, Germany (click here for a short bio)
6 Dec 2021
Title: Benchmarking artificial intelligence methods for end-to-end computational pathology
Abstract: Artificial intelligence (AI) can extract subtle visual information from digitised histopathology slides and yield biological insight on genotype-phenotype interactions as well as clinically actionable recommendations. In the last years, diverse development of the computational techniques resulted in advanced algorithms like Multiple-Instance learning (MIL) and clustering-constrained attention MIL (CLAM), which their performances have been evaluated on specific and clinically relevant targets. However, it is unclear how these different approaches perform relative to each other. We performed a large-scale benchmarking study and implemented and systematically compared five methods in six clinically relevant end-to-end prediction tasks. In particular, we compared a classical residual neural network (ResNet), a modern convolutional neural network (EfficientNet), and non-convolutional vision transformers (ViT) with classical MIL and CLAM. We showed that histological tumour sub-typing of renal cell carcinoma is an easy task which all approaches can successfully solve with an area under the receiver operating curve (AUROC) of above 0.9 without any significant differences between approaches. In contrast, we report significant performance differences for mutation prediction in colorectal, gastric, and bladder cancer. Weakly supervised ResNet- and ViT-based workflows significantly outperformed other methods, in particular MIL and CLAM for mutation prediction. This benchmark study provides important actionable advice for future studies and real-world applications of computational pathology.
Dr Çiğdem Gündüz Demir
Koç University, Istanbul, Turkey
20 Dec 2021
Radboud University Medical Center, The Netherlands (click here for a short bio)
17 Jan 2022