부산대학교 도서관

Primary brain tumors are among the ten most common causes of cancer-related death, and gliomas are the most prevalent type of adult brain tumor, accounting for 78% of malignant brain tumors. Accurate glioma subtype classification is critical to treating patients with brain tumors; however, developing an automated computer-aided diagnosis system for glioma is not trivial. This study introduces the development of an AI-assisted diagnosis system for adult glioma. There are four subtypes of glioma targeted, including Oligodendroglioma (O), Anaplastic Oligodendroglioma (AO), Astrocytoma (A), and Glioblastoma (G). Instead of directly classifying all whole-slide images into four categories, the hierarchical framework with a multi-view scheme was proposed. That is, two-stage predictions were performed. The first stage is to classify all slides into the Oligo-like and Astro-like classes by observing cells' morphologies with high magnification (x40) views by the patching approach. Subsequently, the second stage is to quantify the areas of normal vessels, hyperplasia vessels (vascular proliferation), and necrosis regions with views of lower magnification (x5, x10) by YOLO segmentation, predicting grading levels for each type. After false detection reduction, the experimental results show the quantification of vessels and necrosis regions are the valuable features to indicate grading levels, and the performance of the proposed two-stage hierarchical framework in the final 4-class predictions surpassed the approach of classifying all slides into 4 categories directly.