Work presented at the Ethics and Explainability for Responsible Data Science (EE-RDS) Conference 2021.
[ Video recording of presentation ] | [ Slides ] | [ Paper ]
The internal behavior of Deep Neural Network architectures can be difficult to interpret. Certain architectures achieve impressive feats in a particular dataset while failing to show comparable performance in other datasets. Developing an architecture that performs well on a dataset can be a time-consuming affair and computationally intensive process. This study explains the effect of transfer learning by fine-tuning already available state-of-the-art architectures in different datasets and using them to classify Chest X-Ray images with high accuracy. Using transfer learning helps the model learn problem-specific features in a short period. It further shows that different models perform differently in a particular setting for a dataset. Ablation studies show that a combination of smaller structures that gives an overall better result may not give the best result in the combined model. In addition, the “belief” of the model for selecting a particular class is visualized in this study.
https://figshare.com/articles/dataset/CXR_Challenge_QUBIQ_21_COVID19_Perc_Estimation/21225287
Note: (While using the model, use these labels)
index = {'normal': 0, 'covid': 1, 'pneumonia': 2}
rev_index = {0: 'normal',1: 'covid', 2: 'pneumonia'}
The authors are thankful to Moulay Akhloufi for sharing the datasets.
The authors are also grateful to Swathy Prabhu Mj, Ramakrishna Mission Vivekananda Educational and Research Institute, for arranging a machine with an Asus RTX 2080 Ti (12 GB VRAM) and 64 GB RAM, to hasten the research.
@INPROCEEDINGS{9708580,
author={Pal, Jimut Bahan and Paul, Nilayan},
booktitle={2021 Ethics and Explainability for Responsible Data Science (EE-RDS)},
title={Classifying Chest X-Ray COVID-19 images via Transfer Learning},
year={2021},
volume={},
number={},
pages={1-8},
doi={10.1109/EE-RDS53766.2021.9708580}
}
@misc{CovidGrandChallenge2021,
author = {Akhloufi, Moulay A. and Chetoui, Mohamed},
title = {{Chest XR COVID-19 detection}},
howpublished = {\url{https://cxr-covid19.grand-challenge.org/}},
month = {August},
year = {2021},
note = {Online; accessed September 2021},
}