Optic disc segmentation model improved by contextual information and attention mechanism(PDF)
《中国医学物理学杂志》[ISSN:1005-202X/CN:44-1351/R]
- Issue:
- 2023年第1期
- Page:
- 47-53
- Research Field:
- 医学影像物理
- Publishing date:
Info
- Title:
- Optic disc segmentation model improved by contextual information and attention mechanism
- Author(s):
- WANG Zhenhua; LIU Yangxing; ZHAO Xiaoyu; ZHANG Shutai
- School of Information, Shanghai Ocean University, Shanghai 201306, China
- Keywords:
- Keywords: optic disc segmentation glaucoma contextual information attention mechanism Transformer
- PACS:
- R318；TP391.41
- DOI:
- DOI:10.3969/j.issn.1005-202X.2023.01.008
- Abstract:
- Abstract: Glaucoma which is one of the main causes of blindness is a multiple fundus disease. Considering that fundus images come from a wide range of sources and vary in quality, and that the optic disc region has multi-scale nature, contextual information is beneficial to accurately segmenting the multi-scale optic disc. Based on U-Net, an improved optic disc segmentation model is proposed by combining contextual information and convolutional block attention module (CBAM). The backbone network (ResNet34) is improved by attention mechanism and instance-batch normalization module for enhancing the generalization of segmentation model and the ability to extract image channel features. A multi-level context extraction (MCE) module is proposed for processing the features output from the backbone network and enhancing the ability of the segmentation model to extract the edge features the optic disc. The extraction of the multi-scale features of the optic disc and image channel features is further improved by replacing skip connections and up-sampling with Transformer mechanism. The optic disc segmentation performance of the proposed model is compared with different segmentation models, such as U-Net, U-Net++, DeeplabV3+, FCN and PSPNet. The results show that the proposed segmentation model has better optic disc segmentation results, and achieves Dice, MIoU, MPA and FPS of 98.18%, 96.45%, 98.11% and 17.56 Img/s respectively. The study can provide technical support for the early diagnosis of glaucoma.
Last Update: 2023-01-07