Deep learning-based survival prediction for multiple cancer types using histopathology images
Problem
Predicting disease-specific survival risk for cancer patients at diagnosis from histopathology images, where existing staging and clinical variables leave room for better risk stratification.
Why it matters
Prognostic information at diagnosis has important implications for treatment and monitoring decisions.
What the AI did
A weakly-supervised deep learning system predicted disease-specific survival across 10 TCGA cancer types from histopathology slides without pixel-level annotations, testing three survival loss functions.
Results
Hazard ratio 1.58 (95% CI 1.28-1.70, p<0.0001) for disease-specific survival in the combined 10-cancer cohort after adjusting for cancer type, stage, age, and sex; developed on 9,086 slides from 3,664 cases, evaluated on 3,009 slides from 1,216 cases.
Impact
Peer-reviewed DLS significantly associated with survival beyond standard clinical variables across 10 cancer types.
Record
| Venue | PLoS ONE |
|---|---|
| Published | 2020 |
| Work type | article |
| Authors | Ellery Wulczyn, David F. Steiner, Zhaoyang Xu, Apaar Sadhwani, Hongwu Wang, Isabelle Flament-Auvigne, Craig H. Mermel, Po-Hsuan Cameron Chen et al. |
| Citations | 265 (OpenAlex, at import time) |
| References | 37 |
| Model or method | weakly-supervised deep learning system; specific architecture not stated in source |
| Organization | not stated in source |
| Primary source | https://doi.org/10.1371/journal.pone.0233678 |
| Open access | gold |
| Keywords | survival prediction, histopathology, deep learning, TCGA, weak supervision, Cox regression |
Importance rating
significant — thousands to millions affected
Prognostic risk stratification across 10 cancer types beyond staging; treatment/monitoring implications stated, but no population scale given.