Hepatitis B Surface Antigen Intensity Assessment from Immunohistochemical Stained Images Using Hierarchical Nucleus-guided Cytoplasm Grading Method (Preprint)

Abstract

Previous studies have shown that the likelihood of liver cancer recurrence scales proportionally with the hepatitis B surface antigen (HBsAg) intensity. The intensity assessment of HBsAg is traditionally performed manually based on an observation of the cytoplasm in stained immunohistochemistry (IHC) whole slide images (WSIs). However, the assessment process is time consuming and subjective, and depends heavily on the experience of the pathologist. To address these issues, the use of machine learning to provide objective results has been suggested as a potential solution. However, applying deep learning networks to perform supervised cytoplasmic segmentation is challenging due to the lack of clear boundaries in the cytoplasm, which makes it difficult to generate accurate and a large number of ground truth data for training. This presents a significant challenge for HBsAg intensity assessment.

This study presents the first automatic approach to grade the HBsAg intensity from IHC stained WSIs. The proposed method combines supervised learning with unsupervised techniques to provide quantitative results without the need of detailed and comprehensive labeling of cytoplasmic boundaries. The HBsAg intensity assessment enables pathologists to objectively evaluate the relationship between HBsAg and liver cancer by eliminating the need for human intervention.

A hierarchical nucleus-guided cytoplasm grading method is proposed for HBsAg intensity assessment. The proposed method comprises three sequential steps: nuclear instance segmentation, unsupervised cytoplasm region extraction, and HBsAg intensity grading. The output of the nuclear segmentation network is used to guide an unsupervised cytoplasm region extraction process. HBsAg intensity grading is then performed in the HSV color domain to obtain quantitative analysis results. The proposed method leverages the power of convolutional neural networks (CNNs) to facilitate the challenging task of unsupervised cytoplasm region extraction and enable the automatic HBsAg intensity assessment.

The feasibility of the proposed method is demonstrated using an NCKUH HBsAg intensity grading dataset consisting of 40 WSIs acquired from patients with clinically-diagnosed liver disease. The results show that a grading accuracy of 95% is obtained for the testing images.

The proposed method provides a novel automatic approach for assessing the HBsAg intensity from IHC stained images and visualizing the intensity assessment results in a user-friendly and intuitive manner. It thus provides a convenient and objective assessment tool to support pathologists in performing HBsAg intensity grading and analyzing the correlation between HBsAg and liver cancer.