Detecting COVID-19-infected regions in Lung CT scan through a novel dual-path Swin Transformer-based network

Abstract

Background: Deep learning-based automatic segmentation provides significant advantages over traditional manual segmentation methods in medical imaging. Current approaches for segmenting regions of Coronavirus disease 2019  (COVID-19) infections mainly utilize convolutional neural networks (CNNs), which are limited by their restricted receptive fields (RFs) and consequently struggle to establish global context connections. This limitation negatively impacts their performance in accurately detecting complex details and boundary patterns within medical images.

Methods: This study introduces a novel dual-path Swin Transformer-based network to address these limitations and enhance segmentation accuracy. Our proposed model extracts more informative 3D input patches to capture long-range dependencies and represents both large and small-scale features through a dual-branch encoder. Furthermore, it integrates features from the two paths via the new Transformer Interactive Fusion (TIF) module. The architecture also incorporates an inductive bias by including a residual convolution (Res-conv) block within the encoder.

Results: The proposed network has been evaluated using a 5-fold cross-validation technique, alongside data augmentation, on the publicly available COVID-19-CT-Seg and MosMed datasets. The model achieved Dice coefficients of 0.872 and 0.713 for the COVID-19-CT-Seg and MosMed datasets, respectively, demonstrating its effectiveness relative to prior methodologies.

Conclusion: The significant improvements in segmentation accuracy, demonstrated by the achieved Dice coefficients on the COVID-19-CT-Seg and MosMed datasets, highlight the potential of our approach to enhance automated segmentation in medical imaging.