DEEP-COVID-SEV: AN ENSEMBLE 2D AND 3D CNN-BASED APPROACH FOR COVID-19 SEVERITY PREDICTION FROM 3D CT-SCANS

Since the advent of Covid-19 in late 2019, medical image analysis with artificial intelligence (AI) has become an important research topic. CT-scan imaging is an important diagnostic tool for this disease. This study is part of the 3rd COV19D competition for Covid-19 Severity Prediction, where we aim to close the significant gap between validation and testing results of the previous competition. To achieve this, we proposed two methods based on 2D and 3D CNN, respectively. Our 2D-CNN approach, called 2B-InceptResnet, includes two paths for segmented lungs and for infection of all slices of the input CT-scan, respectively. Each path contains a ConvLayer and an Inception-ResNet model pre-trained on ImageNet. In contrast, our 3D-CNN approach, known as Hybrid-DeCoVNet, consists of four blocks: Stem, four 3D-ResNet layers, classification head, and decision layer. Decision-based ensemble models are also created using these two proposed solutions with six training subsets. Our proposed approaches outperformed the baseline approach by 36% in the validation data of the 3rd COV19D competition for predicting the severity of Covid-19. In addition, our approach ranked second in the testing phase with an improvement of 14% compared to the baseline approach. These promising results demonstrate the potential of our novel method to improve the diagnosis and prognosis of Covid-19, which could contribute to the development of better treatment strategies and ultimately save lives.