Skip to main content

Lund University Publications

LUND UNIVERSITY LIBRARIES

LGANet : Local-Global Augmentation Network for Skin Lesion Segmentation

Guo, Qingqing ; Fang, Xianyong ; Wang, Linbo ; Zhang, Enming LU and Liu, Zhengyi (2023) 20th IEEE International Symposium on Biomedical Imaging, ISBI 2023 In Proceedings - International Symposium on Biomedical Imaging 2023-April.
Abstract

Automatic segmentation of skin lesion is still challenging due to ambiguous boundary and noise interference of lesion regions. Recent exiting Transformer-based methods often directly apply Transformer to obtain long-range dependency to overcome these problems. However, they generally do not consider that patch partitioning strategy of Transformer could lead to the loss of local details around boundaries. Furthermore, dependencies across local windows only represent global information at a coarse level. Therefore, to overcome the limitations, two novel modules, Local Focus Module (LFM) and Global Augmentation Module (GAM) are proposed in this paper. LFM learns the local context around boundary regions to strengthen the discrimination... (More)

Automatic segmentation of skin lesion is still challenging due to ambiguous boundary and noise interference of lesion regions. Recent exiting Transformer-based methods often directly apply Transformer to obtain long-range dependency to overcome these problems. However, they generally do not consider that patch partitioning strategy of Transformer could lead to the loss of local details around boundaries. Furthermore, dependencies across local windows only represent global information at a coarse level. Therefore, to overcome the limitations, two novel modules, Local Focus Module (LFM) and Global Augmentation Module (GAM) are proposed in this paper. LFM learns the local context around boundary regions to strengthen the discrimination between classes. And GAM learns the global context at a finer level to enhance global feature representation. Integrating LFM and GAM, a new Transformer encoder based framework, Local-Global Augmentation Network (LGANet), is proposed. LGANet is efficient in segmenting lesions with ambiguous boundary and with noise interference and its performances are demonstrated with extensive experiments on two public skin lesion segmentation datasets.

(Less)
Please use this url to cite or link to this publication:
author
; ; ; and
organization
publishing date
type
Chapter in Book/Report/Conference proceeding
publication status
published
subject
keywords
Global dependency, Local detail information, Skin lesion segmentation, Transformer
host publication
2023 IEEE International Symposium on Biomedical Imaging, ISBI 2023
series title
Proceedings - International Symposium on Biomedical Imaging
volume
2023-April
publisher
IEEE Computer Society
conference name
20th IEEE International Symposium on Biomedical Imaging, ISBI 2023
conference location
Cartagena, Colombia
conference dates
2023-04-18 - 2023-04-21
external identifiers
  • scopus:85172141985
ISSN
1945-8452
1945-7928
ISBN
9781665473583
DOI
10.1109/ISBI53787.2023.10230358
language
English
LU publication?
yes
id
6f824e6d-9f3f-4cc6-af12-af9f1df851be
date added to LUP
2023-12-20 14:06:34
date last changed
2024-04-29 15:39:18
@inproceedings{6f824e6d-9f3f-4cc6-af12-af9f1df851be,
  abstract     = {{<p>Automatic segmentation of skin lesion is still challenging due to ambiguous boundary and noise interference of lesion regions. Recent exiting Transformer-based methods often directly apply Transformer to obtain long-range dependency to overcome these problems. However, they generally do not consider that patch partitioning strategy of Transformer could lead to the loss of local details around boundaries. Furthermore, dependencies across local windows only represent global information at a coarse level. Therefore, to overcome the limitations, two novel modules, Local Focus Module (LFM) and Global Augmentation Module (GAM) are proposed in this paper. LFM learns the local context around boundary regions to strengthen the discrimination between classes. And GAM learns the global context at a finer level to enhance global feature representation. Integrating LFM and GAM, a new Transformer encoder based framework, Local-Global Augmentation Network (LGANet), is proposed. LGANet is efficient in segmenting lesions with ambiguous boundary and with noise interference and its performances are demonstrated with extensive experiments on two public skin lesion segmentation datasets.</p>}},
  author       = {{Guo, Qingqing and Fang, Xianyong and Wang, Linbo and Zhang, Enming and Liu, Zhengyi}},
  booktitle    = {{2023 IEEE International Symposium on Biomedical Imaging, ISBI 2023}},
  isbn         = {{9781665473583}},
  issn         = {{1945-8452}},
  keywords     = {{Global dependency; Local detail information; Skin lesion segmentation; Transformer}},
  language     = {{eng}},
  publisher    = {{IEEE Computer Society}},
  series       = {{Proceedings - International Symposium on Biomedical Imaging}},
  title        = {{LGANet : Local-Global Augmentation Network for Skin Lesion Segmentation}},
  url          = {{http://dx.doi.org/10.1109/ISBI53787.2023.10230358}},
  doi          = {{10.1109/ISBI53787.2023.10230358}},
  volume       = {{2023-April}},
  year         = {{2023}},
}