DeepTree: Pathological Image Classification Through Imitating Tree-Like Strategies of Pathologists

Area of Science:

• Computational pathology
• Artificial intelligence in medicine
• Digital pathology

Background:

• Digitization of pathological slides enables AI-driven computer-aided diagnosis.
• Current AI methods in computational pathology often overlook pathologists' prior knowledge and diagnostic reasoning, particularly regarding lesion morphology.
• There is a need for AI tools that integrate pathological expertise for more accurate and interpretable results.

Purpose of the Study:

• To introduce DeepTree, a novel deep learning architecture inspired by pathologists' diagnostic decision-making processes.
• To enhance the accuracy, transparency, and clinical applicability of AI in pathological image analysis.
• To leverage prior knowledge of diagnostic strategies for superior representation in computational pathology.

Main Methods:

• Designed a novel deep learning architecture, DeepTree, employing a binary tree structure to mimic pathological diagnosis.
• Developed a tree-like strategy that conditionally learns correlations between tissue morphology.
• Optimized network branches to refine performance and validated the approach on lung cancer and breast tumor datasets.

Main Results:

• DeepTree demonstrated improved accuracy, transparency, and convincing results in pathological image classification.
• The tree-like strategy incorporating prior knowledge showed superior representation ability compared to existing methods.
• The methodology enhanced pathologist trust in AI analysis and promoted practical clinical applications.

Conclusions:

• DeepTree offers a more accurate, transparent, and trustworthy AI-driven approach for pathological image analysis.
• Integrating prior diagnostic knowledge into AI models significantly improves their representation capabilities.
• This work facilitates the clinical adoption of AI-assisted pathological diagnosis by aligning with expert reasoning.