Visual Evaluative AI - EvaSKan

Visual Evaluative AI is a tool for decision support by providing positive and negative evidence for a given hypothesis. This tool finds high-level human concepts in an image and generates the Weight of Evidence (WoE) for each hypothesis in the decision-making process. We also apply this tool in the skin cancer domain by building a web-based application that allows users to upload a dermatoscopic image, select a hypothesis and analyse their decisions by evaluating the provided evidence.

By applying this tool, we build a web-based application called EvaSKan to evaluate dermatoscopic images. Users can select a hypothesis and the application will generate positive/negative evidence for that particular hypothesis.

Prerequisites

Environment

Python 3.10.4
CUDA 12.2.0
UCX-CUDA 1.13.1-CUDA-12.2.0
cuDNN 8.9.3.28-CUDA-12.2.0
Graphviz 5.0.0
torch 2.1.2+cu121
torchvision 0.16.2+cu121

Installation

virtualenv ~/venvs/venv-3.10.4-cuda12.2 #create the env
source ~/venvs/venv-3.10.4-cuda12.2/bin/activate #activate the env
pip3 install torch torchvision torchaudio
pip3 install -r requirements.txt

Code structure

datasets
├── 7-point-criteria
└── HAM10000
save_model
EvaluativeAI
├── Explainers
├── README.md
├── app.py
├── classifiers.py
├── config.py
├── eval.py
├── ice
│   ├── __init__.py
│   ├── channel_reducer.py
│   ├── explainer.py
│   ├── model_wrapper.py
│   └── utils.py
├── ice_clfs.py
├── learn_concepts_dataset.py
├── main.py
├── pcbm
│   ├── concepts
│   │   ├── __init__.py
│   │   └── concept_utils.py
│   ├── data
│   │   ├── __init__.py
│   │   ├── concept_loaders.py
│   │   ├── constants.py
│   │   └── derma_data.py
│   └── models
│       ├── __init__.py
│       ├── derma_models.py
│       ├── model_zoo.py
│       └── pcbm_utils.py
├── pcbm_output
├── preprocessing
│   ├── cnn_backbones.py
│   ├── data_utils.py
│   ├── initdata.py
│   └── params.py
├── reproducibility
│   ├── output
│   └── script
│       ├── pretrained.sh
│       └── scratch.sh
├── requirements.txt
├── test_data
├── train_cnn.py
├── utils.py
└── woe
    ├── __init__.py
    ├── explainers.py
    ├── utils.py
    ├── woe.py
    └── woe_utils.py

Datasets

HAM10000 dataset: Skin lesion classification dataset - Kaggle Link
Derm7pt: Dermatology concepts dataset - Link

Please put the datasets in the right folder followed the code structure above.

Usage

Reproducibility

To reproduce the results in the paper Visual Evaluative AI: A Hypothesis-Driven Tool with Concept-Based Explanations and Weight of Evidence, please either use pre-trained models or train from scratch as described below. Then, run python eval.py to see the results.

Use pre-trained models

Available pre-trained models Download
- Pre-trained CNN backbones Resnet50, ResneXt50 and Resnet152: in the folder save_model
- Pre-trained concept models ICE, PCBM: pickle files in the folder Explainers
- Pre-trained concept bank for PCBM: in the folder pcbm_output
Please refer to reproducibility/script/pretrained.sh for the training using pre-trained models above

Train from scratch

Step by step to train from scratch
- Train the CNN backbone model
- For unsupervised learning concept, train the concept model ICE
- For supervised learning concept, we first need to train the concept bank using the 7pt checklist dataset, then train the concept model PCBM using the HAM10000 dataset
Please refer to reproducibility/script/scratch.sh for training from scratch

Run the app

Before running the app, please download the for_app.zip Download and follow the code structure to put them in the right place.

# Use an unsupervised learning concept model (ICE)
python app.py --algo ice

# Use a supervised learning concept model (PCBM)
python app.py --algo pcbm

References

Citation

@misc{le2024visualevaluativeaihypothesisdriven,
      title={Visual Evaluative AI: A Hypothesis-Driven Tool with Concept-Based Explanations and Weight of Evidence},
      author={Thao Le, Tim Miller, Ruihan Zhang, Liz Sonenberg, Ronal Singh},
      year={2024},
      eprint={2407.04710},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2407.04710},
}

thaole25 / EvaluativeAI