Multiscale Fusion and Boundary Refinement UNet Model for Lung Parenchyma Segmentation

Huang Hehe, Wei Senyu, Zhao Chen, Shi Yu, Zhao Yuetao, Li Yuan, Zang Qiuming, Xu Jianhua
Proceedings of 2024 International Conference on Machine Learning and Intelligent Computing, PMLR 245:57-66, 2024.

Abstract

Methods based on automatic computer-aided systems for lung cancer diagnosis have gained popu-larity in recent years. Lung parenchyma segmentation technology plays an important role in these methods. To reduce the gradient loss, improve the feature utilization, and alleviate the difficulty of fully mining the contextual information in lung parenchyma segmentation, this study proposes a network model for lung parenchyma segmentation based on a Multiscale fusion and Boundary Refinement UNet (MBR–UNet) model. First, along the encoding path, the features are efficiently extracted by a Residual–Residual block module. Next, along the decoding path, a multiscale at-tention–spatial pyramid pooling module fully integrates the feature maps of different layers and sums the outputs of each layer of the decoding path for boundary refinement. Finally, the training model is optimized through a hybrid loss function. The proposed model is experimentally evaluated on the lung segmentation dataset of the Kaggle competition. The accuracy, Dice similarity coefficient, intersection ratio, and Hausdorff distance of the network segmentation are improved by 98.79%, 97.35%, 96.34%, and 12.82 mm, respectively, from those of other segmentation methods. According to these results, the method can more precisely segment pulmonary parenchyma than the existing methods.

Cite this Paper

BibTeX
@InProceedings{pmlr-v245-hehe24a, title = {Multiscale Fusion and Boundary Refinement UNet Model for Lung Parenchyma Segmentation}, author = {Hehe, Huang and Senyu, Wei and Chen, Zhao and Yu, Shi and Yuetao, Zhao and Yuan, Li and Qiuming, Zang and Jianhua, Xu}, booktitle = {Proceedings of 2024 International Conference on Machine Learning and Intelligent Computing}, pages = {57--66}, year = {2024}, editor = {Nianyin, Zeng and Pachori, Ram Bilas}, volume = {245}, series = {Proceedings of Machine Learning Research}, month = {26--28 Apr}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v245/main/assets/hehe24a/hehe24a.pdf}, url = {https://proceedings.mlr.press/v245/hehe24a.html}, abstract = {Methods based on automatic computer-aided systems for lung cancer diagnosis have gained popu-larity in recent years. Lung parenchyma segmentation technology plays an important role in these methods. To reduce the gradient loss, improve the feature utilization, and alleviate the difficulty of fully mining the contextual information in lung parenchyma segmentation, this study proposes a network model for lung parenchyma segmentation based on a Multiscale fusion and Boundary Refinement UNet (MBR–UNet) model. First, along the encoding path, the features are efficiently extracted by a Residual–Residual block module. Next, along the decoding path, a multiscale at-tention–spatial pyramid pooling module fully integrates the feature maps of different layers and sums the outputs of each layer of the decoding path for boundary refinement. Finally, the training model is optimized through a hybrid loss function. The proposed model is experimentally evaluated on the lung segmentation dataset of the Kaggle competition. The accuracy, Dice similarity coefficient, intersection ratio, and Hausdorff distance of the network segmentation are improved by 98.79%, 97.35%, 96.34%, and 12.82 mm, respectively, from those of other segmentation methods. According to these results, the method can more precisely segment pulmonary parenchyma than the existing methods. } }
Endnote
%0 Conference Paper %T Multiscale Fusion and Boundary Refinement UNet Model for Lung Parenchyma Segmentation %A Huang Hehe %A Wei Senyu %A Zhao Chen %A Shi Yu %A Zhao Yuetao %A Li Yuan %A Zang Qiuming %A Xu Jianhua %B Proceedings of 2024 International Conference on Machine Learning and Intelligent Computing %C Proceedings of Machine Learning Research %D 2024 %E Zeng Nianyin %E Ram Bilas Pachori %F pmlr-v245-hehe24a %I PMLR %P 57--66 %U https://proceedings.mlr.press/v245/hehe24a.html %V 245 %X Methods based on automatic computer-aided systems for lung cancer diagnosis have gained popu-larity in recent years. Lung parenchyma segmentation technology plays an important role in these methods. To reduce the gradient loss, improve the feature utilization, and alleviate the difficulty of fully mining the contextual information in lung parenchyma segmentation, this study proposes a network model for lung parenchyma segmentation based on a Multiscale fusion and Boundary Refinement UNet (MBR–UNet) model. First, along the encoding path, the features are efficiently extracted by a Residual–Residual block module. Next, along the decoding path, a multiscale at-tention–spatial pyramid pooling module fully integrates the feature maps of different layers and sums the outputs of each layer of the decoding path for boundary refinement. Finally, the training model is optimized through a hybrid loss function. The proposed model is experimentally evaluated on the lung segmentation dataset of the Kaggle competition. The accuracy, Dice similarity coefficient, intersection ratio, and Hausdorff distance of the network segmentation are improved by 98.79%, 97.35%, 96.34%, and 12.82 mm, respectively, from those of other segmentation methods. According to these results, the method can more precisely segment pulmonary parenchyma than the existing methods.
APA
Hehe, H., Senyu, W., Chen, Z., Yu, S., Yuetao, Z., Yuan, L., Qiuming, Z. & Jianhua, X.. (2024). Multiscale Fusion and Boundary Refinement UNet Model for Lung Parenchyma Segmentation. Proceedings of 2024 International Conference on Machine Learning and Intelligent Computing, in Proceedings of Machine Learning Research 245:57-66 Available from https://proceedings.mlr.press/v245/hehe24a.html.

Related Material