Automated Segmentation of COVID-19 Lung Infections From Chest CT Images via 2.5-D Dual-Path Encoder With Multiscale Dynamic Weight Deep Supervision Mechanism

Accurate segmentation of coronavirus disease 2019 (COVID-19)-infected lung areas from chest computed tomography (CT) scans is critical for COVID-19 diagnosis and prognosis evaluations. However, current mainstream 2-D slice-wise segmentation models fail to fully utilize 3-D volumetric context across slices, leading to inaccurate segmentation results. Although 3-D segmentation models can exploit volumetric information, they are prone to overfitting due to the limited open datasets of COVID-19 and the lack of clinical expert labels. To address the above issues, we propose a novel COVID-19 lesion segmentation model, called 2.5-D DPE-Net, based on a 2.5-D dual-path encoder network model with a multiscale dynamic weight deep supervision mechanism (MsDWDSM). MsDWDSM guides layer-by-layer refinement of multiscale feature maps, enabling the model to learn hierarchical representations of features at different scales to improve the explicitness of the hidden layer learning process. The 2.5-D DPE-Net can fuse spatial and interslice features to effectively enhance and convey spatial location information about the lesion and improve segmentation accuracy. By combining two public datasets, we conducted extensive confirmatory and comparative experiments on a multivendor, multicenter COVID-19 CT image segmentation dataset. Our experimental results show that the 2.5-D DPE-Net performs significantly better than the state-of-the-art methods, with sensitivity, specificity, precision, and Dice of 88.57%, 99.84%, 68.55%, and 75.89%, respectively.