Extracting interpretable features for pathologists using weakly supervised learning to predict p16 expression in...

Area of Science:

• Oncology
• Pathology
• Artificial Intelligence

Background:

• Existing artificial intelligence (AI) models for histopathology lack interpretability.
• Accurate prediction of p16-positive oropharyngeal squamous cell carcinoma (OPSCC) is crucial for diagnosis and treatment.

Purpose of the Study:

• To develop an interpretable AI model for extracting and analyzing p16-positive OPSCC features.
• To enhance the understanding of histopathological characteristics associated with p16 expression in OPSCC.

Main Methods:

• A clustering-constrained attention-based multiple-instance learning (CLAM) model was developed using whole-slide images from 114 OPSCC cases.
• Tumor annotation was incorporated (Annot-CLAM) to improve model performance, achieving a mean AUC of 0.905.
• Histopathologic morphological analysis and CycleGAN image translation were used to examine AI-identified image patches.

Main Results:

• Significant differences in nuclear number, perimeter, and intercellular bridges were observed between p16-negative and p16-positive image patches.
• CycleGAN-converted images confirmed significant alterations in nuclear size and density.
• The AI model successfully identified interpretable histopathological features relevant to p16 status.

Conclusions:

• This novel AI approach significantly improves the interpretability of histopathology-based AI models.
• The study advances the identification of clinically valuable histopathological features for OPSCC.
• The findings contribute to more accurate and understandable AI-driven cancer diagnostics.