Archetype Based Conformal Intervals (ABCI)

This repository provides the official implementation of the methods presented in our paper, Reliable and Interpretable Visual Field Progression Prediction with Diffusion Models and Conformal Risk Control, which has been accepted for publication at MICCAI 2025.

Installation

1. Install the required dependencies:

   cd puq
   pip install -e .

📥 Download Dataset

To run the experiments, download the required datasets from the links below:

• UW Dataset: Download
• GRN Dataset: Download

unzip them in the 'puq/data' folder.

💌 Basic Usage

To run the script with default settings:

python run.py --no-cache --archetypes --alpha 0.25 

🏥 UW Dataset Experiments

🟢 Mild Case (q = 0.95, β = 0.1)

python run.py --no-cache --archetypes --data data/UW_subgroups/mild --alpha 0.25 --q 0.95 --beta 0.1 

🟡 Moderate Case (q = 0.9, β = 0.14)

python run.py --no-cache --archetypes --data data/UW_subgroups/moderate --alpha 0.25 --q 0.9 --beta 0.14

🔴 Severe Case (q = 0.9, β = 0.155)

python run.py --no-cache --archetypes --dataset data/UW_subgroups/severe --alpha 0.25 --q 0.9 --beta 0.155

---

🧠 GRN Dataset Experiments

🟢 Mild Case (q = 0.95, β = 0.1)

python run.py --no-cache --archetypes --data data/Scheie_subgroups/mild --alpha 0.25 --q 0.95 --beta 0.1

🟡 Moderate Case (q = 0.9, β = 0.15)

python run.py --no-cache --archetypes --data data/Scheie_subgroups/moderate --alpha 0.25 --q 0.9 --beta 0.15

🔴 Severe Case (q = 0.9, β = 0.2)

python run.py --no-cache --archetypes --data data/Scheie_subgroups/severe --alpha 0.25 --q 0.9 --beta 0.2

🚀 Additional Notes

• --no-cache: Ensures the script runs fresh without using cached results.
• --archetypes: Enables the archetypal analysis module.
• --dataset: Specifies the dataset (UW or GRN).
• --q and --beta: Control the model's sensitivity and regularization.