GastricCancer-NAIT-Radiopathomics

Preoperative Prediction of pCR in Gastric Cancer (Chapter 3)

This repository contains the source code for the research: "Preoperative Prediction of Pathological Complete Response (pCR) in Gastric Cancer based on CT Radiomics and Biopsy Pathology".

📑 Research Overview

This study proposes a multi-modal fusion framework to predict immunotherapy response (pCR) in gastric cancer patients. By integrating CT imaging, biopsy whole-slide imaging (WSI), and standardized clinical text, we achieve superior predictive performance compared to single-modality models.

📁 Repository Structure (Based on Chapter 3)

Following the organization in the dissertation, the code is structured as follows:

• Data_process/: LLM-driven text standardization and WSI preprocessing.
• Models/: Implementation of the AB-MIL (Attention-based Multi-instance Learning) and Fusion architectures.
• Engine/: Training pipelines and cross-validation logic.
• Configs/: Hyperparameter settings for different modalities.
• utils/: Common tools for data loading and feature extraction.
• script.py: The main entry point for model execution.

🛠 Clinical Evaluation Tools

The repository includes scripts for generating critical clinical metrics discussed in Section 3:

• ROC Curves: Performance comparison across modalities.
• Calibration Analysis: Reliability check for probability predictions.
• Decision Curve Analysis (DCA): Clinical utility and net benefit assessment.

🚀 Key Technologies

• AB-MIL Aggregation: Necessary for biopsy scenarios with limited samples.
• Multi-modal Fusion: Progressive gain effect through CT and Pathology integration.
• Standardized NLP: LLM-driven preprocessing for medical text.

✉️ Contact

For review purposes, this repository is currently anonymized. For technical inquiries, please refer to the contact information provided in the dissertation.