Integrated tumor and microenvironment subtyping refines the colorectal cancer classification system with prognostic and therapeutic implications

This repository contains the official implementation of CT4CMSPlus, a deep learning framework for predicting CMSPlus molecular sutypes of colorectal cancer from preoperative CT scans.

Installation

---

• Python 3.8

• PyTorch 2.4.1, torchvision 0.19.1

• CUDA 11.3

• For common packages, simply run:

  pip install -r requirements.txt

  If additional dependencies are missing, please install them manually.

Data Preprocessing

---

All preprocessing scripts are located in the preprocess_data folder.

1. Colon CT Preprocessing

This script preprocesses colon CT scans by performing:

• Spatial alignment validation between CT and segmentation masks.
• Anisotropic resampling to 1.0 × 1.0 × 3.0 mm spacing.
• Hounsfield Unit (HU) windowing normalization with WL = 35, WW = 300.

python preprocess_colon_ct.py

2. Radiomics Feature Extraction

This unified script performs two major operations:

• Radiomic Feature Extraction

• Significant Radiomics Features Selection

python pyradiomics/extract_and_select_radiomics_features.py

3. Dataset Configuration

Datasets should be defined in structured JSON format for model training and inference.

• Example — training_dataset.json

{
  "fold0": {
    "training": {
      "images/image1.nii.gz": {
        "label": 0,
        "roi": "labels/image1_roi.nii.gz"
      },
      "images/image2.nii.gz": {
        "label": 1,
        "roi": "labels/image2_roi.nii.gz"
      }
    },
    "validation": {
      "images/image3.nii.gz": {
        "label": 0,
        "roi": "labels/image3_roi.nii.gz"
      },
      "images/image4.nii.gz": {
        "label": 1,
        "roi": "labels/image4_roi.nii.gz"
      }
    }
  },
  "fold1": { "...": "..." }
}

• Example — inference_dataset.json

{
  "training": null,
  "validation": {
    "images/image1.nii.gz": {
      "label": "CMS4",
      "roi": "labels/image1_roi.nii.gz"
    },
    "images/image2.nii.gz": {
      "label": "CMS4",
      "roi": "labels/image2_roi.nii.gz"
    }
    ...
  }
}

CT4CMSPlus Model

---

All inference scripts are located in the train_infer_code folder.

Inference

1. CMS Classification (CT4CMS)

Use CT4CMS to first predict CMS1–CMS4 molecular subtypes from preoperative CT scans:

Reference: [arXiv: Preprint link]

2. Fine Subclassification (CT4CMSPlus)

Then, use CT4CMSPlus to further stratify CMS4 cases into CMS4-TME(+) and CMS4-TME(-) subtypes.

Before running inference, prepare inference_dataset.json.

python inference.py --data_path ./preprocessed/1_1_3  \
                    --label_path ./path/inference_dataset.json \
                    --radiomics_path ./path/selected_radiomics_features.csv \
                    --pretrained ./pretrained_weights/model.pth \
                    --optimal_threshold 0.5

Acknowledgements

---

We thank the developers of:

• [CT4CMS] for CMS classification
• Med3D (MedicalNet) for 3D-ResNet pretrained models used for transfer learning

License

---

This project is licensed for academic research use only.

Commercial or clinical usage is prohibited without explicit permission.