Transfer Learning with Limited Data for COVID-19 Chest X-Ray Classification

Overview

This project demonstrates the application of transfer learning in limited data scenarios, in the medical field. Data collection in medical fields has been scarce and difficult because of variations in patient data and imaging techniques, in addition to legal restrictions and many other reasons, so the use of transfer learning to analyze patterns even with very limited data is crucial.

To demonstrate, we are going to use the COVID-19 Chest X-ray Dataset, shrunk down to contain only 150 training images. We focus on the importance of choosing the correct model for transfer learning which varies based on the type of data and the application, as in it is case specific.

Objectives

1. To perform and observe the performance of transfer learning on a small COVID-19 Chest X-ray Dataset.
2. Some pre-trained models comparison to find the best performer.

Dataset

The dataset used here has chest X-ray images to classify whether a patient is suffering from COVID-19 or not, in addition to viral pneumonia which can be caused by Covid-19 but is difficult to differentiate in x-ray scans.

Models Used with Transfer Learning

Pre-trained models tried:

• ResNet50
• VGG16
• InceptionV3
• MobileNetV2

All these models were trained on imagenet. Those models were attached to a classifier and fine-tuned on the dataset for further improving their performance.

Key Observations

• The variance of model performances indicates that the choice of model is very critical, specially when dealing with limited data.

• DenseNet and VGG16 were the best performers on our dataset, while being very simple to fine tune.

• ResNet50 was the worst performer, although it was not fine-tuned due to the complexity of tuning it.

• If a model does not perform well on this dataset, that does not mean the model is bad, it simply means that the dataset is not suitable for the model.

Conclusion

This work thus shows that transfer learning has endless applications and is crucial to creating a simpler development environment for AI progress. Experimenting with multiple models when choosing a model is extremely important for better results, as well as providing a varying level of knowledge needed to work with the model as some are more complex to work with.