Attention-Based Color Spatial Relationship Matrix for Optimizing Deep Learning Performance in Histopathology Image Grayscale Conversion
5
Issued Date
2025-01-01
Resource Type
Scopus ID
2-s2.0-105003426516
Journal Title
Proceedings - 2025 5th Asia Conference on Information Engineering, ACIE 2025
Start Page
80
End Page
85
Rights Holder(s)
SCOPUS
Bibliographic Citation
Proceedings - 2025 5th Asia Conference on Information Engineering, ACIE 2025 (2025) , 80-85
Suggested Citation
Srisermphoak N., Amornphimoltham P., Chaisuparat R., Achararit P., Fuangrod T. Attention-Based Color Spatial Relationship Matrix for Optimizing Deep Learning Performance in Histopathology Image Grayscale Conversion. Proceedings - 2025 5th Asia Conference on Information Engineering, ACIE 2025 (2025) , 80-85. 85. doi:10.1109/ACIE64499.2025.00020 Retrieved from: https://repository.li.mahidol.ac.th/handle/123456789/109906
Title
Attention-Based Color Spatial Relationship Matrix for Optimizing Deep Learning Performance in Histopathology Image Grayscale Conversion
Author's Affiliation
Corresponding Author(s)
Other Contributor(s)
Abstract
Whole Slide Imaging (WSI) is a crucial image format in Digital Pathology, particularly for malignant diagnosis. It facilitates the application of Deep Learning (DL) to lighten pathologist's workloads in manual tasks. However, the format's size poses a challenge for DL clinical adoption, as it requires high bandwidth and a stable connection to successfully transfer to a DL operating source. By accounting for the characteristics of Hematoxylin and Eosin (H&E) that represent tissue structure with limited color shades, the utilization of grayscale conversion can aid the challenge. However, conventional grayscale methods are unsuitable for these tasks because they emphasize unessential characteristics for histopathology image analysis. Accordingly, this study proposes the Attention-based Color Spatial Relationship Matrix (ACSRM), a grayscale conversion approach inspired by the Transformer's Attention mechanism. ACSRM captures long-range relationships of every pixel across RGB channels, ensuring grayscale outputs retain critical color information to maintain DL performance as trained on color images and simultaneously reducing file size by threefold. Evaluations on histopathological classification and segmentation datasets demonstrate ACSRM's potential to preserve color information for DL. When applied to Convolutional Neural Network (CNN)-based architectures, it presents equivalent-level performance against the color input (i.e., the RGB format) on every task. These findings highlight ACSRM as a scalable and practical alternative to colorbased inputs in DL workflows for histopathology image analysis tasks due to its capability to balance efficiency and accuracy.