Official code for the paper: Melanoma classification from dermatoscopy images using knowledge distillation for highly imbalanced data published in Computers in Biology and Medicine, Elsevier. A live demo is available here.

TL;DR Melanoma classification task is challenging due to the high inter-class and low intra-class similarity problems in dermoscopic image datasets. The work proposes a novel knowledge-distilled lightweight Deep-CNN-based framework to tackle the high inter-class and low intra-class similarity problems with Knowledge Distillation, Cost-Sensitive Learning with Focal Loss for addressing class imbalance to achieve better sensitivity scores.

Citation

If you use this code in your work, please cite the following paper:

@article{adepu2023melanoma,
  title={Melanoma classification from dermatoscopy images using knowledge distillation for highly imbalanced data},
  author={Adepu, Anil Kumar and Sahayam, Subin and Jayaraman, Umarani and Arramraju, Rashmika},
  journal={Computers in Biology and Medicine},
  pages={106571},
  year={2023},
  publisher={Elsevier}
}

Proposed framework for Melanoma Classification using Knowledge Distillation

Major Requirements

• Tensorflow: 2.4.0 or above
• TensorFlow Addons: 0.14.0 or above
• Python: 3.7 or above

Processed Datasets

• JPEG: https://www.kaggle.com/datasets/akra98/isic2020-jpg-256x256-inpainted2. Code to re-generate these images is available under preprocess directory above.
• TFRECORDS: https://www.kaggle.com/datasets/aniladepu/isic2020-tfrec-256x256-inpainted2. A sample code to create tfrecords is available here.