Awesome

DUE

This repo is for DUE: Dynamic Uncertainty-Aware Explanation Supervision via 3D Imputation

Setup

Please refer to requirements.txt

Implementation

1. Train a diffusion model for slice interpolation
2. Interpolate annotation slices repeatly using the diffusion model to estimate uncertainty
3. Train a VAE to directly estimate uncertainty
4. Set paths and run:
   python main.py --model "due" --dataset $DATASET --attention_weight 1 --seed 0

Deployment

Below are screenshots illustrating the deployment of DUE, using lung nodule classification as an example:

1. Visual Annotation Labeling Interface

The screenshots below display the interface for labeling visual annotations. Radiologists can annotate images by drawing on them, generating a binary matrix of the focus area. This process contributes to enhancing the quality of model explanations.

2. Model Selection Interface

Here is the interface for selecting the model, where users can choose from trained model checkpoints.