Interpretable Graph Convolutional Networks with CPC features for Whole Slide Histology Classification
Principal Investigator
Name
Faisal Mahmood
Degrees
Ph.D.
Institution
Brigham and Women's Hospital
Position Title
Assistant Professor
Email
About this CDAS Project
Study
NLST
(Learn more about this study)
Project ID
NLST-592
Initial CDAS Request Approval
Oct 29, 2019
Title
Interpretable Graph Convolutional Networks with CPC features for Whole Slide Histology Classification
Summary
Lung cancer is the leading cause of cancer death in the world, accounting for 2.1 million new cases and 1.8 million deaths in 2018. Due to late diagnosis and lack of early treatment interventions, lung cancer has a poor 5-year relative survival rate, ranging from 6%-60% at the first state of diagnosis. Most current deep learning-based histology classification methods are limited to extracted ROIs and require pixel-level annotation. Moreover, such methods are not data-efficient and sufficiently context-aware i.e. they do not make use of the explicit spatial relationships between cells. This project aims to develop data-efficient methods for whole slide analysis using contrastive predictive coding (CPC) and graph convolutional networks (GCNs) that do not require slide level labels.
Aims
Specific Aims
Aim 1: Developing graph convolutional networks for capturing the spatial contiguity and structure from histology images for classification.
Aim 2: Using contrastive predictive coding (CPC) to improve GCN performance for whole slide level classification without pixel-level annotations.
Aim 3: Using the developed methods for ROI detection within the TCGA lung cancer cohort as a validation of the methods developed.
Collaborators
Max Lu, Harvard Medical School and Brigham and Women's Hospital
Richard Chen, Harvard Medical School and Brigham and Women's Hospital
Jingwen Wang, Harvard Medical School and Brigham and Women's Hospital