Kudos to the amazing team at @Path_AI for making this happen! Shima Nofallah, Maryam Pouryahya, Janani Iyer, @KenLeidal, John Abel, @alibehrooz (8/8)

If you are interested in this line of research, join us at our talk at the workshop on Computer Vision for Microscopy Image Analysis on June 19th at 2:52 PM CT. (7/8)

Beyond NASH, this paper provides a general method for incorporating multiple modalities of image information into a single inference without the need to align the images beforehand–critical for working on multi-stain problems in pathology. (6/8)

In NASH, we found that inferring scores from fused H&E and Trichrome graphs yielded upto 20% better concordance with a panel of pathologists. Importantly, we show our results on data from 6 clinical trials, demonstrating viability of our approach in real-world conditions. (5/8)

Our proposed GNN fusion model aggregates human interpretable CNN features via graph representations of non-registered whole slide images of differently-stained tissue specimens. This model extracts and integrates complementary multimodal information for disease staging. (4/8)

When pathologists assess slides, such as when grading and staging liver tissue in non-alcoholic steatohepatitis (NASH), they view tissue samples from the same patient with different stains to reveal complementary information about the diseased tissue. (3/8)

In computational pathology, it’s more common to infer on single slides rather than a set of slides that may uniquely characterize a disease. To overcome limitations of the one-model-per-modality structure, we present a single fusion model that leverages multiple slides. (2/8)

Excited to be presenting our recent work “Multi stain graph fusion for multimodal integration in pathology” at CVPR 2022. 🧵 Paper: openaccess.thecvf.com/conten… #CVPR2022 #ComputationalPathology #ArtificialIntelligence (1/8)