An Efficient Honeycomb Lung Segmentation Network Combining Multi-Paradigms Representation and Cascade Attention

Honeycomb lung is a pulmonary manifestation that occurs in the terminal stage of various lung diseases, which greatly threatens patients. Due to the different locations and irregular shapes of lesions, the accurate segmentation of the honeycomb region is an essential and challenging problem. However, most deep learning methods struggle to effectively utilize both global and local information from lesion images, resulting in cannot to accurately segment the lesion. In addition, these methods often ignore some semantic information that is necessary for the segmentation of lesion location and shape in the decoding stage. To alleviate these challenges, in this paper, we propose a dual-branch encoder and cascaded decoder network (DECDNet) for segmenting honeycombs lesions. First, we design a dual-branch encoder consisting of ResNet34 and Swin-Transformer with different paradigm representations to extract local features and long-range dependencies respectively. Next, to further combine the different paradigm features, we develop the feature fusion module to obtain richer representation information. Finally, considering the problem of information loss during the decoder, a cascaded attention decoder is constructed to aggregate the multi-stage encoder information to get the final segmentation result. Experimental results demonstrate that our method outperforms other methods on the in-house honeycomb lung dataset. Notably, compared with the other nine universal methods, the proposed DECDNet obtains the highest IoU (86.34%), Dice (92.66%), Precision (93.21%), Recall (92.13%), F1-Score (92.66%), and achieves the lowest HD95 (7.33) and ASD (2.30). In particular, our method enables precisely segmenting lesions under different clinical scenarios as well. Our code and dataset are available at https://github.com/ybq17/DECDNet.

@article{Yang2023,
  title     = {An Efficient Honeycomb Lung Segmentation Network Combining Multi-Paradigms Representation and Cascade Attention},
  journal   = {International Journal of Advanced Computer Science and Applications},
  volume    = {14},
  number    = {12},
  year      = {2023},
  publisher = {The Science and Information Organization},
  author    = {Bingqian Yang and Xiufang Feng and Yunyun Dong},
  doi       = {10.14569/IJACSA.2023.0141256},
  url       = {https://doi.org/10.14569/IJACSA.2023.0141256}
}