SI-ViT: Shuffle instance-based Vision Transformer for pancreatic cancer ROSE image classification

Area of Science:

• Computational pathology
• Artificial intelligence in medical diagnostics
• Image analysis and machine learning

Background:

• Rapid On-site Evaluation (ROSE) aids pancreatic cancer diagnosis using fast-stained cytopathological images.
• Automating ROSE classification faces challenges from staining variations, device differences, and complex cancerous patterns.
• Variability in cell patterns complicates precise identification and relationship modeling in cytopathological samples.

Purpose of the Study:

• To develop an automated classification method for rapid on-site evaluation (ROSE) of pancreatic cancer.
• To address challenges in ROSE classification caused by image perturbations and sample variability.
• To enhance the diagnostic capabilities of pathologists through an AI-driven approach.

Main Methods:

• Proposed an instance-aware approach, Shuffle Instance Vision Transformer (SI-ViT), enhancing the Vision Transformer (ViT).
• Introduced a novel shuffle instance strategy involving a shuffle step to create instance bags and soft-labels.
• Incorporated an un-shuffle step to enable traditional ViT to model sample label relationships, focusing on inner-sample and cross-sample instance relationships.

Main Results:

• Achieved significant improvements in ROSE classification compared to state-of-the-art methods.
• SI-ViT demonstrated interpretable attention regions for identifying cancerous and normal cells.
• Validated generalization potential on diverse pathological image datasets.

Conclusions:

• Instance relationship modeling via shuffling offers a novel approach to pathological image analysis.
• Significant improvements in ROSE classification suggest potential for AI-on-site applications in pancreatic cancer diagnosis.
• The developed method enhances accuracy and interpretability in automated cytopathological analysis.