Related Experiment Video
Updated: Dec 14, 2025
04:48
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
3.2K
Image-based consensus molecular subtype (imCMS) classification of colorectal cancer using deep learning
Korsuk Sirinukunwattana1,2,3, Enric Domingo4, Susan D Richman5
1Institute of Biomedical Engineering (IBME), Department of Engineering Science, University of Oxford, Oxford, UK.
Gut
|July 22, 2020
Summary
Deep learning on H&E images accurately predicts colorectal cancer molecular subtypes (CMS). This image-based CMS approach offers a cost-effective method for cancer stratification, improving routine diagnostic workflows.
Area of Science:
- Computational pathology
- Digital pathology
- Cancer genomics
Background:
- Histological features in colorectal cancer (CRC) lack clear links to molecular subtypes.
- Current methods for CRC molecular classification require gene expression profiling, limiting accessibility.
- Histological grading is an unreliable predictor of CRC progression.
Purpose of the Study:
- To develop and validate an image-based deep learning model for predicting CRC Consensus Molecular Subtypes (CMS) from standard H&E stained slides.
- To assess the cost-effectiveness and reliability of image analysis for cancer stratification.
- To explore the potential of image analysis in resolving intratumoural heterogeneity and unclassifiable cases.
Main Methods:
- A deep learning neural network was trained on 1206 H&E tissue sections with multi-omic data from three independent cohorts (FOCUS, GRAMPIAN, TCGA).
- Ground truth CMS calls were established using random forest and single-sample prediction classifiers.
- The model's performance was evaluated on unseen datasets, including TCGA and rectal cancer biopsies.
Main Results:
- The image-based CMS (imCMS) model achieved high classification accuracy on independent datasets (AUC=0.84 for TCGA, AUC=0.85 for rectal cancer).
- imCMS demonstrated the ability to spatially resolve intratumoural heterogeneity and provided secondary classifications correlating with molecular data.
- The model successfully classified previously unclassifiable samples and showed prognostic associations similar to transcriptomic CMS.
Conclusions:
- Histopathological images analyzed with deep learning can predict RNA expression-based CRC molecular subtypes.
- This image-based approach provides a simple, cost-effective, and reliable method for biological stratification in routine pathology workflows.
- imCMS has the potential to enhance cancer subtyping and improve patient stratification for treatment.
