Segmentation of Plain CT Image of Ischemic Lesion based on Trans-Swin-UNet

Abstract

The present study aims to build a hybrid convolutional neural network and transformer UNet-based model, Trans-Swin-UNet, to segment ischemic lesions of the plain computed tomography (CT) image. The model architecture is built based on TransUnet and has four main improvements. First, replace the decoder of TransUNet with a Swin transformer; second, add a Max Attention module into the skip connection; third, design a comprehensive loss function; and last, speed up the segmentation performance. The present study designs two experiments to evaluate the performance of the built model using both the self-collected and public plain CT image datasets. The model optimization experiment evaluates the improvements of Trans-Swin-UNet over TransUnet. The experimental results show that each improvement of the built model can achieve a better performance than TransUNet in terms of dice similarity coefficient (DSC), Jaccard coefficient (JAC), and accuracy (ACC). The comparison experiment compares the built model with four existing UNet-based models. The experimental results show that the built model had a DSC of 0.72±0.01, a JAC of 0.78±0.04, an ACC of 0.75±0.03 using the self-collected plain CT image dataset and a DSC of 0.73±0.02, a JAC of 0.79±0.03, an ACC of 0.76±0.02 using the public plain CT image dataset, achieving the best segmentation performance among five UNet-based neural network models. The two experimental results conclude that the built model could accurately segment ischemic lesions of the plain CT image. The limitations and future work of this study are also discussed.