Deep-Adaptation: Ensembling and Test Augmentation for Covid-19 Detection and Covid-19 Domain Adaptation from 3D CT-Scans

Fares Bougourzi, Feryal Windal Moulai, Halim Benhabiles, Fadi Dornaika, Abdelmalik Taleb-Ahmed; Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2024, pp. 5097-5104

Since the onset of the Covid-19 pandemic in late 2019 the realm of medical image analysis has seen a surge in importance particularly with the utilization of CT-scan imaging for disease diagnosis. This paper presents findings from our participation in the 4th COV19D competition specifically targeting the challenges of Covid-19 Detection and Covid-19 Domain Adaptation. Our methodology revolves around lung segmentation and Covid-19 infection segmentation employing the state-of-the-art CNN-based segmentation architecture PDAtt-Unet. Unlike conventional methods we introduce a novel approach by concatenating the input slice (grayscale) with segmented lung and infection thereby generating three input channels reminiscent of color channels. Furthermore we leverage three distinct 3D CNN backbones--Customized Hybrid-DeCoVNet in addition to pretrained 3D-Resnet-18 and 3D-Resnet-50 models--to facilitate Covid-19 recognition for both challenges. To further boost performance we explore ensemble techniques and testing augmentation. Comparison with baseline results highlights the substantial efficiency of our approach showcasing a significant improvement in terms of F1-score (14%) on the validation data. Our approach ranked second and third in the Covid-19 Detection and Covid-19 Domain Adaptation Challenges respectively based on the test data results. Our approach demonstrates improvements of 9.5% and 17% compared to baseline performance in these challenges. Furthermore our approach exhibits very promising performance compared with the approaches of other competitors underscoring the significance of the proposed training paradigm and the utilization of ensemble and testing augmentation techniques.