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

Area of Science:

• Medical image analysis
• Pathology
• Computer vision

Background:

• Accurate segmentation of lesion tissues in medical microscopic hyperspectral pathological images (MHSIs) is vital for early tumor diagnosis and patient prognosis.
• Complex tissue structures and indistinct boundaries pose significant challenges for precise MHSIs segmentation.

Purpose of the Study:

• To introduce BE-Net, a novel method for precise segmentation of lesion tissues in MHSIs.
• To address the challenges of complex structures and indistinct boundaries in MHSIs segmentation.

Main Methods:

• BE-Net utilizes a multi-scale strategy and edge operators for fine edge detail capture.
• Information entropy is incorporated into attention mechanisms to enhance relevant feature representation.
• Novel components include a Laplacian of Gaussian operator convolution boundary feature extraction block, a grouped multi-scale edge feature extraction module, and a multi-scale spatial boundary feature extraction block.

Main Results:

• BE-Net was evaluated on MHSIs datasets of gastric intraepithelial neoplasia and gastric mucosal intestinal metaplasia.
• Experimental results show BE-Net outperforms state-of-the-art methods in segmentation accuracy and boundary preservation.

Conclusions:

• BE-Net offers a significant advancement in MHSIs segmentation.
• The proposed method improves accuracy and boundary preservation, aiding in early tumor diagnosis.