Cross-Scale Hypergraph Neural Networks with Inter-Intra Constraints for Mitosis Detection.

Area of Science:

• Pathology
• Artificial Intelligence
• Computational Biology

Background:

• Mitotic figures in tumor tissues are crucial for diagnosing malignant lesions.
• Manual inspection of whole slide imaging (WSI) for mitosis is time-consuming and prone to human error.
• Current AI diagnostic models face challenges with high computational costs, suboptimal accuracy, and lack of intercellular information.

Purpose of the Study:

• To develop an advanced AI model for accurate and efficient detection of mitotic figures in tumor tissues.
• To address the limitations of existing models by incorporating intercellular context.
• To improve cytopathological diagnosis through deep learning and hypergraph neural networks.

Main Methods:

• Proposed an Inter-Intra Hypergraph Neural Network (II-HGNN) model.
• Implemented a block-based feature extraction mechanism for efficient deep representation.
• Utilized hypergraph convolutional networks to model both intracellular and intercellular information.

Main Results:

• The II-HGNN model demonstrated superior performance compared to baseline models.
• The approach achieved higher accuracy in detecting mitotic figures.
• Consistent outperformance across varying imaging conditions on public datasets.

Conclusions:

• The II-HGNN model effectively captures both intracellular and intercellular information for precise cytopathological diagnosis.
• This AI-driven approach offers a promising solution for enhancing the accuracy and efficiency of tumor diagnosis.
• The model's ability to consider cellular context represents a significant advancement in AI-assisted pathology.