Project Description:
Histopathology has significantly contributed to the understanding of biological phenomena and many diseases. It typically involves visual evaluation of a tissue sample under a light microscope by pathologists to identify structural tissue properties associated with diseases. The emerging use of Whole-Slide Imaging (WSI) in Digital Pathology at large-scale/high-throughput is associated with a number of novel scientific challenges including: vast amounts of large images; variance of signal, i.e. intra-stain variance, and staining, i.e. inter-stain variance; and images that are highly heterogeneous. These cause difficulties when applying conventional image processing algorithms. To date, robustness to cross-centre analyses is not sufficiently solved. HistoGraph brings together three computer science laboratories, two specialised in AI, machine learning, and medical image analysis and one specialised in machine learning with graph representations, which will work in collaboration with medical institutes to overcome this problem.
This consortium will develop AI based diagnostic approaches by WSI analysis with aim:
- (1) to segment multiple anatomical structures and cells using deep learning approaches in WSIs originating from multiple-sites (hospitals) without additional annotations;
- (2) create graph-based approaches to naturally capture the spatial context of the segmented objects and allow multi-modal information to be integrated;
- (3) be interpretable to extract additional information for the diagnostic approach and for users;
- (4) rigorous evaluation using standard datasets, and potential clinical application in collaboration with pathologist.
This will be achieved through the following 5 work packages, each jointly carried out by two laboratories:
- WP1 - Data Collection
- WP2 - Segmentation & Domain Invariance
- WP3 - Graph Modelling
- WP4 - Interpretability
- WP5 - Diagnosis Approach & Evaluation
Thus, WP1–3 will be supported by ICube and IHU; and WP4 and WP5 by l’X and IHU.