Multi-scale attention based CNN algorithm for Covid-19 classification

Abstract

Coronavirus 2019 (COVID-19) is a new acute respiratory disease that has spread rapidly throughout the world. The main purpose of this project is to investigate and compare several different deep learning enhanced neural network techniques applied to CT-scan medical images for the detection of COVID-19. In this project, we used some powerful neural network models with deep learning enhancement, such as classical CNN, RNN, LSTM, Multi-scale CNN and Multi-scale CNN with Attention mechanism, for the COVID-19 CT-scan binary classification task. The proposed project framework is designed to find out the best architecture, pre-processing, and training parameters for the models largely automatically. The accuracy and F1-score were both above 98% and 91% respectively, in the diagnosis of COVID-19 using CT-scan images. In addition, we applied augmentation techniques to overcome the insufficient data and to improve the training time. The accuracy and validity of the algorithms were assessed on CT-scan images of well-known public datasets. The proposed method has better results for COVID-19 diagnosis than others related in literature. In our opinion, our work can help virologists and radiologists to make a better and faster diagnosis in the struggle against the outbreak of COVID-19. This project proposes a lightweight or novel convolutional neural network (CNN) model named Multi-Scale Attention based CNN (AMS-CNN) is proposed, which is based on attention mechanism and multi-scale convolution. A multi-scale attention mechanism is designed to extract effective feature information from the COVID-19 CT images for classification. Moreover, the multiscale convolution layers are adopted as AMS-CNN’s backbone to reduce the model size and parameters. The training, verification/ validation and testing are carried out on an interesting CT scan images dataset in Kaggle named ‘sarscov2-ctscan-dataset’. The AMSCNN model is compared with classical CNN, RNN, LSTM and Multi-scale CNN models. The experimental results show that the MS-CNN model with three multi-scale convolution layers with Attention mechanism (AMSCNN) has achieved accuracy of 98.02% on CT scan images dataset, which is higher than the other four models. The specificity and sensitivity are 90.71% and 92.18% on CT scan images dataset. The size of the AMSCNN model is only 57.28 MB, and in addition, the detection speed of the model on CT scan images is 4.23 ms per image, respectively. It is proved that our AMSCNN model can detect and classify COVID-19 faster with higher accuracy and efficiency.