Method

The goal of the model is to use resnet50 with AMP(AUTOMATIC MIXED PRECISION) to accelerate the training speed and make a high-resolution skin cancer classification.

Usage

crop_transform.py - self-definition val_aug method. Transform 1 test/val image to 9 images to predict the score.

dataset2017.py - method to read your ISIC Training / Val / Test Data.

generate_patch_images.ipynb - Because the augmentation strategy contains a random method. Please use this method to transform your Training Data from 2000 images to 122000 images first.

amp related: predict2017_amp_bce_testaug_avg9score_v.ipynb train_resnet50_amp_v.ipynb original: predict2017_woamp_bce_testaug_avg9score_v.ipynb train_resnet50_woamp_v.ipynb

Release Note

* v1.0.0, 2023/08/05, 15:45:00

Citation

Paper: J. Zhang, Y. Xie, Y. Xia and C. Shen, "Attention Residual Learning for Skin Lesion Classification," in IEEE Transactions on Medical Imaging, vol. 38, no. 9, pp. 2092-2103, Sept. 2019, doi: 10.1109/TMI.2019.2893944.

Original third-party code: https://github.com/Vipermdl/ARL

Acknowledgements

This work was supported in part by the National Science and Technology Council, Taiwan under Grant NSTC 111-2634-F-006-012.
We thank National Center for High-performance Computing (NCHC) for providing storage resources.