Hierarchically Optimized Multiple Instance Learning With Multi-Magnification Pathological Images for Cerebral Tumor...

Area of Science:

• Computational pathology
• Artificial intelligence in oncology
• Machine learning for medical diagnosis

Background:

• Accurate cerebral tumor diagnosis is vital for treatment and prognosis.
• Limitations in biopsy tissue and pathologist expertise hinder precise diagnosis.
• Developing automated diagnostic tools is essential for neuro-oncology.

Purpose of the Study:

• To develop a novel computational method for accurate brain tumor diagnosis.
• To classify six common brain tumor types, grade gliomas, and determine metastatic origins.
• To overcome limitations of traditional diagnostic methods using machine learning.

Main Methods:

• Established a large brain tumor dataset (3,520 cases) from multiple centers.
• Developed a Hierarchically Optimized Multiple Instance Learning (HOMIL) method.
• Alternately trained feature encoder and aggregator based on specific datasets and tasks.

Main Results:

• HOMIL achieved state-of-the-art performance on classification, grading, and origin determination tasks.
• Achieved high accuracies: 93.29% (classification), 91.21% (glioma grading), 86.36% (origin determination) on internal datasets.
• Effectively located regions of interest and visualized critical areas on pathological slides.

Conclusions:

• HOMIL offers a robust and accurate solution for complex brain tumor diagnostics.
• The method demonstrates superior performance compared to existing multiple instance learning approaches.
• HOMIL facilitates in-depth analysis and aids in identifying critical diagnostic regions on pathology slides.