COVLIAS 3.0 XEDL : Multicentre, Cloud-Based, Explainable Ensemble Artificial Intelligence Deep Learning System for COVID- 19 in Computed Tomography Scans

Abstract

Abstract Background and Motivation: Lung computed tomography (CT) techniques have been utilized in the intensive care unit (ICU) for COVID-19 disease characterization due to its high-resolution imaging. Artificial Intelligence (AI) has significantly helped researchers in diagnosing COVID-19, and the proposed study hypothesized that the cloud-based explainable ensemble deep learning (XEDL) paradigm is superior to transfer learning (TL) models for disease classification. Methodology: We propose a cloud-based ensemble deep learning (EDL) approach to classify COVID-19 versus Control patients. In the proposed study two cohorts are used: (i) 80 Croatian COVID-19 and (ii)70 Italian COVID-19 patients and 30 Control Italian patients. ResNet-SegNet-based lung segmentation of CT scans on five different data combinations (DC1-DC5) using two cohorts have been designed. Five deep convolutional neural network models namely, DenseNet-169, DenseNet-121, DenseNet-201, EfficientNet-B1, and EfficientNet-B6 models are utilized for ensemble. The focal loss function is used with a gamma value of 2. Five-fold cross-validation has been performed during model training and testing on unseen data. Statistical analysis and heatmaps are generated to validate the model. This model was also available for global use on Amazon Web Services as COVLIAS 3.0XEDL. The proposed COVLIAS 3.0XEDL is superior to TL models. Results The XEDL showed an accuracy of 99.99%, AUC 1 (p < 0.0001) for DC1, 98.23%, AUC 0.97 (p < 0.0001) for DC5, 96.45%, AUC 0.92 (p < 0.0001) for DC2, 88.20%, AUC 0.85 (p < 0.0001) for DC3, and 87.87%, AUC 0.81 (p < 0.0001) for DC4. The proposed XEDL accuracy was 8.59% superior to the mean TL accuracy. Conclusions Our hypothesis holds true where XEDL is superior to TL in a cloud-based explainable framework using heatmaps.