PdCR

Code for the paper "Leveraging Causal Reasoning Method for Explaining Medical Image Segmentation Models".

---

Project Structure

We provide two usage modes:
main.py is the main evaluation script used in the paper, which computes PdCR maps for all models on the test set.
main_one_image.py is a quick demo for single-image testing. Users can specify any end-to-end model and a custom RoI to visualize the segmentation behavior.

Environment Requirements

This project has been tested and runs successfully under the following environment:

• Python: 3.8.20
• PyTorch: 2.4.1+cu121
• Torchvision: 0.19.1+cu121
• numpy: 1.24.4
• timm: 0.4.12
• ninja: 1.11.1.4
• jinja2: 3.14
• mamba-ssm: 2.2.2 (If you need to run the Vision Mamba)

The PdCR method itself is lightweight and compatible with mainstream PyTorch versions. It is recommended to use a CUDA driver version compatible with the above libraries.

⚠️ Note: Different segmentation models may have additional dependencies.

The environment dependencies vary depending on the segmentation model. We recommend referring to the original papers for each model and finding the corresponding code repositories. Please refer to the original papers and corresponding repositories of each model for specific requirements.

---

main.py

Step 1: Prepare a trained model

We provide dataloaders for HAM10000 and FIVES under seg_model/datasets_.
Download these datasets and place them into your own folders, then set the paths in total_config.py via the data_path field.
Optionally, modify the model-specific config.
Use the following command to train any model from model_zoo:

python seg_model/train.py

This will save two checkpoints in the model folder:

• *_latest.pth: for training resume
• *_best.pth: for final evaluation

Step 2: Prepare intervention blocks

Use the following command to randomly select test images for PdCR evaluation, and crop the remaining images to generate perturbation blocks:

python create_PdCR_test_dataset.py

We also provide a small set of 1000 blocks for demo usage under the intervention_patches/ folder.

Step 3: Run main.py

After training, modify model_name_list to select models to test, and set the folder path that stores intervention images.

We provide pre-selected image lists used in the paper under demo/ as two .txt files. These files contain image names and RoI coordinates (top-left corner). Note that these samples come from the original test folder of dataset.
main.py will read them as the coord_file, but you can also replace them with your own .txt file.

We also include the ablation study version of the paper using the zero_local mode for comparison.

---

main_one_image.py

This script only evaluates a single image, but the logic is much simpler and easier to understand.
We recommend starting by loading a simple UNeXt checkpoint and testing the images provided in the demo/ folder.

Example:

We provide the checkpoint file for UNeXt in the demo/ folder. You can easily use it.

python main_one_image.py \
  --model_ckpt_path demo/unext_HAM10000_best.pth \
  --save_path demo/result_folder \
  --image_stem ISIC_0033556 \
  --top_left_i 72 \
  --top_left_j 88