Towards Concept-based Interpretability of Skin Lesion Diagnosis using Vision-Language Models

Paper accepted at IEEE International Symposium on Biomedical Imaging - ISBI 2024.

Citation

If you use this repository, please cite:

@article{patricio2023towards,
  title={Towards Concept-based Interpretability of Skin Lesion Diagnosis using Vision-Language Models},
  author={Patr{\'\i}cio, Cristiano and Teixeira, Lu{\'\i}s F and Neves, Jo{\~a}o C},
  journal={arXiv preprint arXiv:2311.14339},
  year={2023}
}

1. Download data

PH $^2$ dataset: https://www.fc.up.pt/addi/ph2%20database.html
Derm7pt dataset: https://derm.cs.sfu.ca/Welcome.html
ISIC 2018 (Task 3) dataset: https://challenge.isic-archive.com/data/#2018

Note: You should mask out the original images of each dataset with the available masks (download masks here) in order to reproduce the results of the paper.

2. Training

2.1 Prepare conda environment

Create a new conda environment with the required libraries contained in requirements.txt file:

conda create --name cbi-vlm --file requirements.txt

2.2 Fine-Tune CLIP on Derm7pt and ISIC 2018

Use the configuration file (CLIP/modules/config.py) to adjust settings for training:
- clip_model: choose between {ViT-B/32, ViT-B/16, RN50, RN101, ViT-L/14, RN50x16}
- seed: choose between {0, 42, 84, 168}
- dataset: choose between {'derm7pt', 'ISIC_2018'}
- batch_size: default 32
- image_embedding: set accordingly to dim of each CLIP model
- text_embedding: set accordingly to dim of each CLIP model
- projection_dim: set accordingly to your preference
- path_to_model: path of the trained model
See suplementary document for more details on the architectures chosen.
Change image file paths according to your own file paths in extract_image_embeddings function [CLIP/modules/utils.py].
Run train script [CLIP/train.py]:

python train.py

Run inference script [CLIP/inference.py] (Extract image & text embeddings used for evaluation):

 python inference.py

3. Evaluation

All required dataset splits are available under /data folder.

3.1. PH $^2$ dataset

$k$-fold evaluation:

# CLIP - Baseline
python CLIP/scr_k_fold_evaluate_PH2_Baseline.py

# CLIP - CBM
python CLIP/scr_k_fold_evaluate_PH2_CBM.py

# CLIP - GPT-CBM
python CLIP/scr_k_fold_evaluate_PH2_GPT-CBM.py

# MONET - Baseline
python MONET/scr_k_fold_evaluate_PH2_Baseline.py

# MONET - CBM
python MONET/scr_k_fold_evaluate_PH2_CBM.py

# MONET - GPT-CBM
python MONET/scr_k_fold_evaluate_PH2_GPT-CBM.py

# Each of the above scripts will generate a numpy file with the results. Read the file to analyze the results.

Individual evaluation (jupyter notebooks):

# CLIP - Baseline
CLIP/scr_Baseline_CLIP_PH2.ipynb

# CLIP - CBM
CLIP/scr_CBM_CLIP_PH2.ipynb

# CLIP - GPT-CBM
CLIP/scr_GPT-CBM_CLIP_PH2.ipynb

# MONET - Baseline
MONET/scr_Baseline_MONET.ipynb

# MONET - CBM
MONET/scr_CBM_MONET.ipynb

# MONET GPT-CBM
MONET/scr_GPT-CBM_MONET.ipynb

3.2. Derm7pt dataset

Evaluation over four runs:

# CLIP - Baseline
python CLIP/scr_evaluate_derm7pt_Baseline.py

# CLIP - CBM
python CLIP/scr_evaluate_derm7pt_CBM.py

# CLIP - GPT-CBM
python CLIP/scr_evaluate_derm7pt_GPT_CBM.py


# Each of the above scripts will generate a numpy file with the results. Read the file to analyze the results.

Individual evaluation (jupyter notebooks):

# CLIP - Baseline
CLIP/scr_Baseline_CLIP-derm7pt.ipynb

# CLIP - CBM
CLIP/scr_CBM_CLIP-derm7pt.ipynb

# CLIP - GPT-CBM
CLIP/scr_GPT-CBM_CLIP-derm7pt.ipynb

# MONET - Baseline
MONET/scr_Baseline_MONET.ipynb

# MONET - CBM
MONET/scr_CBM_MONET.ipynb

# MONET GPT-CBM
MONET/scr_GPT-CBM_MONET.ipynb

3.3. ISIC 2018 dataset

Evaluation over four runs:

# CLIP - Baseline
python CLIP/scr_evaluate_ISIC_2018_Baseline.py

# CLIP - CBM
python CLIP/scr_evaluate_ISIC_2018_CBM.py

# CLIP - GPT-CBM
python CLIP/scr_evaluate_ISIC_2018_GPT_CBM.py


# Each of the above scripts will generate a numpy file with the results. Read the file to analyze the results.

Individual evaluation (jupyter notebooks):

# CLIP - Baseline
CLIP/scr_Baseline_CLIP-ISIC_2018.ipynb

# CLIP - CBM
CLIP/scr_CBM_CLIP-ISIC_2018.ipynb

# CLIP - GPT-CBM
CLIP/scr_GPT-CBM_CLIP-ISIC_2018.ipynb

# MONET - Baseline
MONET/scr_Baseline_MONET-ISIC_2018.ipynb

# MONET - CBM
MONET/scr_CBM_MONET.ipynb

# MONET GPT-CBM
MONET/scr_GPT-CBM_MONET.ipynb

[Last update: Mon Feb 19 03:41:45 PM WET 2024]

cristianopatricio / concept-based-interpretability-vlm Goto Github PK

concept-based-interpretability-vlm's Introduction

Towards Concept-based Interpretability of Skin Lesion Diagnosis using Vision-Language Models

Citation

1. Download data

2. Training

2.1 Prepare conda environment

2.2 Fine-Tune CLIP on Derm7pt and ISIC 2018

3. Evaluation

3.1. PH $^2$ dataset

3.2. Derm7pt dataset

3.3. ISIC 2018 dataset

concept-based-interpretability-vlm's People

Contributors

Stargazers

Watchers

concept-based-interpretability-vlm's Issues

Recommend Projects

Recommend Topics

Recommend Org