A multi-class classification algorithm based on hematoxylin-eosin staining for neoadjuvant therapy in rectal cancer:
Yihan Wu1,2, Xiaohua Liu3, Fang Liu4
1School of Medicine, Chongqing University, Chongqing, China.
Peerj
|June 19, 2023
Summary
This study developed a novel multi-class classifier using ResNet and Hematoxylin-Eosin images to predict neoadjuvant therapy (NAT) responses in rectal cancer. The model accurately stratifies patients into three groups, addressing a key clinical need for risk assessment.
Area of Science:
- Oncology
- Computational Pathology
- Artificial Intelligence in Medicine
Background:
- Neoadjuvant therapy (NAT) is crucial for locally advanced rectal cancer.
- Current predictive models for NAT response are limited to binary classifications (pathological complete response - pCR).
- Clinical risk stratification requires multi-class prediction of treatment response (TRG0-3).
Purpose of the Study:
- To develop and validate a multi-class classifier for predicting pathological neoadjuvant therapy response in rectal cancer.
- To stratify patients into three distinct response groups: TRG0 (pCR), TRG1/2 (moderate/minimal response), and TRG3 (poor response).
- To identify key pathological features associated with treatment response using explainable AI techniques.
Main Methods:
- Utilized ResNet (Residual Neural Network) architecture for image analysis.
- Trained a multi-class classifier on Hematoxylin-Eosin (HE) stained pathological images.
- Employed Class Activation Mapping (CAM) to generate visual heatmaps and identify predictive image features.
Main Results:
- Achieved high Area Under the Curve (AUC) values: 0.97 at 40x magnification and 0.89 at 10x magnification.
- Demonstrated strong performance across all classes: TRG0 (precision 0.67, sensitivity 0.67, specificity 0.95), TRG1/2 (precision 0.92, sensitivity 0.86, specificity 0.89), and TRG3 (precision 0.71, sensitivity 0.83, specificity 0.88) at 40x.
- Identified tumor nuclei and tumor-infiltrating lymphocytes as potential predictive features through CAM analysis.
Conclusions:
- This study presents the first multi-class classifier capable of predicting diverse neoadjuvant therapy responses in rectal cancer.
- The developed model offers improved risk stratification beyond binary pCR prediction.
- The findings highlight the potential of AI-driven analysis of pathological images for personalized rectal cancer treatment.
