ABSTRACT
Medical imaging has been a crucial component of COVID-19 clinical procedures. However, CT scans for lung and COVID-19 infection diagnosis are time-consuming and require specific knowledge. In addition, diverse imaging modalities, such as regions of interest (organs or lesions) and picture quality, make the automated segmentation of these infectious fragments difficult. In this paper, we proposed TLU-Net based on an optimized design of U-Net architectures to facilitate successful transfer learning for the segmentation of medical images from a source domain, such as lungs, to a target domain, such as COVID-19 infection. We have investigated the effective transfer learning methodologies for U-Net topologies, hyperparameters, freezing, and multiple models/datasets in the lungs and COVID-19 domains. Due to our improved and lightweight U-Net design, experiments with TLU-Net employing three lung and COVID-19 infection domain datasets increased segmentation accuracy and reduced training time. © 2022 IEEE.