LMFLOSS-A-Hybrid-Loss-For-Imbalanced-Medical-Image-Classification

This repository contains the official implementation of the LMFLoss framework from the paper LMFLOSS: A Hybrid Loss For Imbalanced Medical Image Classification.

How to Run

• Install Dependencies:

  • Please install the required libraries from the 'requirements.txt' file.

• Open the 'config.ini' file.

• Define paths:

  • Firstly, define the path of the dataset files in the [dataset-paths] section. The path to a dataset should include three folders: train, test, and val, each of which contains training, testing, and validation images. Additionally, the images should be categorized into folders for each class in the dataset.
  • For example, a dataset location can be defined as, dataset_name = C:\Users......\images
  • Next, define the path in which the trained models will be saved, in the [model-save-paths] section. There needs to be a single defined path for each dataset.
  • For example, a model save location can be defined as, dataset_name = C:\Users......\models

• Adjusting Hyperparameters:

  • The training hyperparameters such as epochs, learning rate, weight decay and others are defined the [hyperparameters] section of the 'config.ini' file. These hyperparameter can be adjusted.
  • There are also seperate sections dedicated for adjusting the hyperparameters of particular loss functions. Such as, [LDAM parameters], [Focal parameters] and [LMF parameters].

• Training:

  • Before starting training first define the 'dataset', 'model', 'loss' and 'model_save_name' variables in the [general] section of the config file.
  • The 'dataset' name should be the same the defined path variable for that dataset.
  • The available keywords for 'model' and 'loss' variables are given in the config file.
  • The 'model_save_name' variable should be set to the name you wish to save your trained model as.
  • Save the 'config.ini' file once you've finished configuring the general and hyperparameter settings.
  • Finally, you can start training by running the 'train.py' file.

• Testing:

  • First, define the model filename(on which you want perform the test) in the [Load Model] section of the config file.
  • The 'dataset' and 'model' should also be defined accordingly in the [general] section.
  • Finally, simply run the 'test.py' file to run the test. The test will also generate a classification report in the code directory.

References

• If you find our paper and code useful, please cite our paper-

@article{sadi2022lmfloss,
 title={LMFLOSS: A Hybrid Loss For Imbalanced Medical Image Classification},
 author={Sadi, Abu Adnan and Chowdhury, Labib and Jahan, Nursrat and Rafi, Mohammad Newaz Sharif and Chowdhury, Radeya and Khan, Faisal Ahamed and Mohammed, Nabeel},
 journal={arXiv preprint arXiv:2212.12741},
 year={2022}
}

• Also cite the papers this work is inspired from-

@inproceedings{lin2017focal,
title={Focal loss for dense object detection},
author={Lin, Tsung-Yi and Goyal, Priya and Girshick, Ross and He, Kaiming and Doll{\'a}r, Piotr},
booktitle={Proceedings of the IEEE international conference on computer vision},
pages={2980--2988},
year={2017}
}

@article{cao2019learning,
  title={Learning imbalanced datasets with label-distribution-aware margin loss},
  author={Cao, Kaidi and Wei, Colin and Gaidon, Adrien and Arechiga, Nikos and Ma, Tengyu},
  journal={Advances in neural information processing systems},
  volume={32},
  year={2019}
}