End-to-end prognostication in colorectal cancer by deep learning: a retrospective, multicentre study
Xiaofeng Jiang1, Michael Hoffmeister2, Hermann Brenner3
1Else Kroener Fresenius Center for Digital Health, Technical University Dresden, Dresden, Germany; Department of Medicine III, University Hospital Rheinisch-Westfälische Technische Hochschule Aachen, Aachen, Germany.
The Lancet. Digital Health
|December 20, 2023
Summary
A new deep learning model accurately predicts colorectal cancer survival using histopathology slides. This tool stratifies patients into high-risk and low-risk groups, aiding clinical decision-making for better patient care.
Area of Science:
- Oncology
- Artificial Intelligence
- Medical Imaging
Background:
- Accurate prognosis prediction is crucial for personalized colorectal cancer (CRC) treatment.
- Histopathological slides of CRC specimens contain vital prognostic information.
- Existing prognostic algorithms lack multicenter external validation and clinical integration.
Purpose of the Study:
- To develop and externally validate a deep learning (DL) system for prognostic stratification in resected colorectal cancer patients.
- To automatically predict overall survival and cancer-specific survival using histopathology.
- To assess the DL model's performance across diverse patient populations and processing protocols.
Main Methods:
- Retrospective, multicenter study involving CRC tissue samples from Australia, Germany, and the USA.
- Development and validation of an attention-based, self-supervised deep learning model.
- Risk stratification of patients into high-risk and low-risk groups based on DL-predicted scores.
Main Results:
- The DL model trained on 4428 patients demonstrated robust prognostic capability.
- High-risk patients exhibited significantly worse overall survival (HR 4.50) and disease-specific survival (HR 8.35) compared to low-risk patients in internal validation.
- Consistent performance was observed across three large external test sets, with HRs for DSS ranging from 2.23 to 3.08.
- The DL-based risk score proved independent of established clinical prognostic factors.
Conclusions:
- Attention-based self-supervised deep learning offers a reliable method for predicting CRC patient outcomes.
- The developed prognostic tool generalizes across populations and can aid clinical decision-making in CRC management.
- Open-source release of codes and models facilitates further research and clinical adoption.
