Boundary aware microscopic hyperspectral pathology image segmentation network guided by information entropy weight

Xueying Cao ¹ Hongmin Gao ¹ Ting Qin ² Min Zhu ¹ Ping Zhang ^3,4 Peipei Xu ^2,5*

• ¹ Hohai University, Nanjing, China
• ² Department of hematology, Nanjing Drum Tower Hospital Clinical College of Nanjing Medical University, Nanjing, China
• ³ School of Information Management and Engineering, Shanghai University of Finance and Economics, Shanghai, Shanghai Municipality, China
• ⁴ College of International Exchange, Nanjing Normal University of Special Education, Nanjing, China
• ⁵ Department of Hematology, Nanjing Drum Tower Hospital, Affiliated Hospital of Medical School, Nanjing University, Nanjing, China

The accurate segmentation of lesion tissues in microscopic hyperspectral pathological images holds importance for enhancing the accuracy of early tumor diagnosis and improving patient prognosis. However, the complex structure and boundaries of lesion tissues create significant challenges in achieving precise segmentation. To address these challenges, we propose a novel method named BE-Net. It employs multi-scale strategy and edge operators to capture fine edge details, while incorporating information entropy to construct attention mechanisms that further strengthen the representation of relevant features. Specifically, we first propose a Laplacian of Gaussian operator-convolution boundary feature extraction block, which encodes feature gradient information through the improved edge detection operators and emphasizes relevant boundary channel weights based on channel information entropy weighting. We further designed a grouped multi-scale edge feature extraction module to optimize the fusion process between the encoder and decoder, with the goal of optimize boundary details and emphasizing relevant channel representations. Finally, we propose a multi-scale spatial boundary feature extraction block to guide the model in emphasizing the most important spatial locations and boundary regions. Experimental results on the microscopic hyperspectral pathological image datasets of gastric intraepithelial neoplasia and gastric mucosal intestinal metaplasia show that BE-Net outperforms other state-of-the-art methods.