A prompt-aware knowledge-tuning framework for histopathology subtype classification with scarce annotation.

Area of Science:

• Computational pathology
• Artificial intelligence in medicine
• Digital histopathology

Background:

• Artificial intelligence (AI) aids pathologists in histopathology subtype diagnosis for precision medicine.
• Current AI models often require extensive annotations and struggle with adaptive feature identification and quantitative knowledge representation.
• Existing methods represent knowledge qualitatively (0 or 1), lacking fine-grained diagnostic insights.

Purpose of the Study:

• To develop a novel AI model, PAKT (prompt-aware knowledge-tuning), for accurate histopathology subtype classification.
• To enable adaptive feature generation and quantitative knowledge representation with scarce annotation.
• To improve diagnostic processes in computational pathology.

Main Methods:

• PAKT utilizes a prompt-aware module for adaptive multi-scale histological probability prediction.
• A pre-trained encoder leverages vision prompts for explicit feature extraction, reducing annotation dependency.
• A knowledge-tuning module with a trainable weight matrix quantitatively represents diagnostic knowledge.

Main Results:

• PAKT demonstrates superior performance over state-of-the-art methods in subtype diagnosis, achieving over 10% average improvement.
• Experiments on public and in-house datasets validate PAKT's effectiveness and robustness.
• The model significantly reduces complexity without compromising performance.

Conclusions:

• PAKT offers an effective solution for histopathology subtype classification using AI.
• The model's ability to adaptively generate features and quantitatively represent knowledge addresses limitations of existing approaches.
• PAKT facilitates precision medicine through improved diagnostic accuracy and efficiency in computational pathology.