Final Project (Gain Full Mark, 40/40)
Deep Transfer Learning for Covid-19 Detection
Recent studies have shown that comprehending and mitigating the low clinical sensitivity of the SARS-CoV-2 RT-PCR testing is pivotal to the containment of the global coronavirus pandemic. In this paper, we propose a complementary method for rapid coronavirus detection from CT scans, based on the deep transfer learning paradigm. The idea is to use fine-tuning methods as a baseline, and to apply mapping-based and generative methods to enhance the transferability of knowledge and to improve the model performance on the Covid-19 detection task. As for the fine-tuning method, the VGG architecture performs the best, achieving an accuracy 75.83%, an F-1 score of 76.68% and an AUC of 83.95%. The mapping-based method is used to predict correct labels while keeping the domain invariant. Applying our selected domain confusion metric on the ResNet50 model, the method achieves an accuracy of 80.00%, an F1-score of 80.93% and an AUC of 86.29%, outperforming the baseline strongly. The generative methods are used to alleviate the training data scarcity issue, associated with the fine-tuning methods. The VGG architecture fine-tuned on the updated dataset with generative and affine augmentations had an accuracy of 78.31%, an F-1 score of 80.45% and an AUC of 85.39%. The results indicate that both proposed methods outperform the baseline and thus bring closer the idea of using deep transfer learning on CT scans as a complementary method for rapid Covid-19 detection.