Skin Lesion Classification and Uncertainty Estimation

This repository includes the TensorFlow implementation of the methods described in our paper Risk-Aware Machine Learning Classifier for Medical Diagnosis.

Fig1. Processing pipeline of the proposed risk-aware Bayesian model.

Demo

A demo of the training and testing with a step-by-step instruction is provided in the Skin_Lesion_Analysis.ipynb file.

Download data and the pretrained model

• Link to get the preprocessed data
• Link to get a sample pretrained model . (create a save folder at .../skin_lesion_uncertainty_estimation/save and paste the lesion_densenet169 model inside it).

Train

• For training with default setup: python main.py

You can easily train your desired network configuration by passing the desired arguments as provided in the config.py file. For example:

• For training with batch size of 8: python main.py --batch_size=8

Test

• For testing the pretrained model run: python inference.py

• For testing your trained model run: python inference.py --model_name=your_model_name_as_saved

Citation

If you found this repo useful, please use this bibtex to cite our paper:

@article{mobiny2019risk,
  title={Risk-Aware Machine Learning Classifier for Skin Lesion Diagnosis},
  author={Mobiny, Aryan and Singh, Aditi and Van Nguyen, Hien},
  journal={Journal of clinical medicine},
  volume={8},
  number={8},
  pages={1241},
  year={2019},
  publisher={Multidisciplinary Digital Publishing Institute}
}