Pyramidal position attention model for histopathological image segmentation. (February 2023)

• Record Type: Journal Article
• Authors: Bozdag, Zehra; Talu, Muhammed Fatih
• Is Part Of: Biomedical signal processing and control. Volume 80:Part 2(2023)
• Publication Date: 2023-02
• Subjects: Deep learning -- Histopathological image segmentation -- Attention module -- Convolution neural network; mIoU mean Intersection over Union -- PA Pixel Accuracy; Signal processing -- Periodicals; Biomedical engineering -- Periodicals; Signal Processing, Computer-Assisted -- Periodicals; Image Processing, Computer-Assisted -- Periodicals; Biomedical Engineering -- Periodicals; 610.28
• DOI: 10.1016/j.bspc.2022.104374
• Languages: English
• ISSNs: 1746-8094

Abstract

Hıghlıghts: A new hybrid network model (PAMSegNet) has been proposed in the histopathological image segmentation task. Attention mechanisms are used differently and more accurate detection of the global information of images is achieved. When the performance comparisons with strong backbone and pre-trained segmentation architectures are evaluated, it is seen that the proposed model can reach high segmentation performance (71.6% mIoU and 86.4% PA) quickly. Abstract: The level of performance achieved in the classification of histopathological images has not yet been reached in the segmentation area. This is because the global context information sufficient for classification is not sufficient for segmentation. Especially, high tissue diversity in histopathological images and the fact that tissues in the same class have quite different colors, patterns and geometries make the segmentation problem difficult. In this study, a novel hybrid architecture (PAMSegNet) is presented that provides high segmentation accuracy in histopathological images. This architecture, which has a pyramid data processing strategy, has been provided with the Position Attention Module (PAM) and Boundary aware Module (BM) to extract global and local attributes more accurately. In addition, with the deep supervised technique used, both contents (global and local) were evaluated together in the segmentation decision. Segmentation architectures (Deeplabv3 +, SegNet, U-Net) with a strong backbone in the literature are used for performance comparison. The proposed architecture has been found to provide high segmentation accuracy (71.6% mIoU and 86.4% PA ).