Abstract

Info

Title:: Pneumonia lesion segmentation method integrating multi-scale feature fusion and attention mechanisms

Author(s):: GUO Yixuan1; GUAN Jiangnan1; WU Jie1; YAO Xufeng2; 1. School of Health Science and Engineering, University of Shanghai for Science and Technology, Shanghai 200093, China 2. School of Medical Imaging, Shanghai University of Medicine and Health Sciences, Shanghai 201318, China

Keywords:: pneumonia lesion image segmentation feature fusion multi-scale mixed attention medical image processing

Abstract:: Abstract: A segmentation model integrating multi-scale features and attention mechanisms is developed to address the challenges of large variations in the size and shape of pneumonia lesions as well as unclear boundaries. The model employs multi-scale visual state space blocks as the main components of both the encoder and decoder. It utilizes cascaded convolution kernels of different sizes to capture multi-scale contextual information from varying-sized lesions, thereby effectively coping with the significant variations in the shape and scale of pneumonia lesions. Meanwhile, a convolutional block attention module is embedded in the skip connection layers. Through a dual-branch interaction mechanism across spatial and channel dimensions, the semantic consistency of encoder-decoder features is improved, and the information loss during feature fusion is alleviated. Experiments are conducted on several public lung pathological image datasets. On the Large COVID-19 CT scan slice dataset and the CNCB pneumonia lesion segmentation dataset, the proposed model achieves Dice similarity coefficients of 87.7% and 85.1%, and IoU of 79.7% and 75.3%, respectively. In addition, the generalization performance of the model is verified on the LUNA16 lung nodule dataset. Experimental results demonstrate that the proposed model exhibits superior segmentation performance over conventional U-Net architectures and Transformer-based models.