.png){kind=icon}
.png){kind=icon}
Return
Polyp Segmentation Based on Multilevel Information Correction Transformer
.png){kind=icon}
OA
AI
DOI：10.1109/ACCESS.2024.3421296.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
This paper proposes a precise and reliable polyp segmentation method using multilevel Information Correction transformer. Unlike the conventional direct fusion approach, it harnesses the strengths of transformers in modeling global information and convolutional neural networks (CNN) in capturing local information. Deep learning methods allow doctors to achieve effective and reliable results, greatly reducing manual costs and improving accuracy. However, previous approaches have not achieved a good balance between incorporating local and global information, leading to segmentation results that are not entirely satisfactory. The proposed method uses an enriched multi-scale feature extractor module(EMFE) that generates reliable boundary attention maps through richer scales. Considering that feature maps at different stages of the encoder have different sizes, the proposed method applies a pyramid hybrid attention module(PHA) to filter out background noise information using different scales. We also devise a multilevel lesion correction module(MLC), utilizing Transformer blocks for target localization, followed by convolutional correction of uncertain regions, effectively harnessing the strengths of both methods to achieve accurate segmentation from coarse to fine. Finally, we introduce a feature selection fusion module(FSF), enabling the network to adaptively select relevant contextual information for further correction, thus achieving an effective balance between global and local information. Experimental results demonstrate significant superiority of our method over the top 17 state-of-the-art approaches on five publicly available datasets. In particular, the proposed method achieves an improvement of 3.36% in dice score and 3.17% in mIoU score compared to the most advanced methods on the ETIS dataset.
Keywords:
Transformers
Task analysis
Convolutional neural networks
Feature extraction
Image segmentation
Accuracy
Decoding
Polyp segmentation
camouflage object detection
long-range dependencies
local details
correction
Journal
IF:
3.6
Papers:
9.8W
Citations:
29.4W