This repository contains the code for the In-Context RCA project. It includes RCA classifier models, a UNet for segmentation, and utilities to support training and inference using PyTorch.

1. Install the required dependencies by running:

   pip install -r requirements.txt

2. Install SAM 2 following the SAM 2 installation instructions.

3. Download the SAM 2 checkpoints and place the folder inside the segment-anything-2.

4. Replace the contents of the sam2 folder with the ones from segment-anything-2. Once the replacement is complete, you may safely delete the segment-anything-2 folder.

• unet_segmentations: Contains different quality segmentations generated with a UNet.
• src: Contains the supported models along with scripts for preprocessing, generating the datasets and evaluation
• scripts: Contains run_inference.py and custom_inference.py scripts for reproducing experiments and running the method on user's custom data.

The following datasets were used:

• SCD
• HC18
• PSFHS
• JSRT
• 3D-IRCAdB
• PH²
• NuCLS
• ISIC 2018
• WBC

To reproduce experiments on the datasets used in the paper, execute the following command with the run_inference.py script:

python run_inference.py --dataset <dataset_path> --classifier <classifier_name> --output_file <output_file_path>

To run inference on your own data, you can use the custom_inference.py script:

python custom_inference.py --ref_dataset <reference_dataset_path> --eval_dataset <eval_data_path> --n_classes <num_of_classes> --classifier <classifier_name> --output_file <output_file_path> 

If you are using our masks please cite our work:

@misc{cosarinsky2025incontextreverseclassificationaccuracy,
      title={In-Context Reverse Classification Accuracy: Efficient Estimation of Segmentation Quality without Ground-Truth}, 
      author={Matias Cosarinsky and Ramiro Billot and Lucas Mansilla and Gabriel Gimenez and Nicolas Gaggión and Guanghui Fu and Enzo Ferrante},
      year={2025},
      eprint={2503.04522},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2503.04522}, 
}