Region of interest (ROI) selection using vision transformer for automatic analysis using whole slide images

Area of Science:

• Digital Pathology
• Medical Image Analysis
• Computational Biology

Background:

• Manual selection of regions of interest (ROI) is critical but time-consuming in whole slide image (WSI) analysis.
• Existing HER2 grading methods are subjective, leading to inter-observer variability and impacting diagnostic accuracy.
• The high resolution and dimensionality of WSI present computational challenges for automated analysis.

Purpose of the Study:

• To develop an automated method for region of interest (ROI) detection in whole slide images (WSI).
• To apply and evaluate the automated ROI detection for human epidermal growth factor receptor 2 (HER2) grading in breast cancer.
• To investigate the impact of image magnification on automated ROI detection performance.

Main Methods:

• Proposed an automated ROI detection method utilizing Vision Transformer architecture.
• Investigated the influence of different image magnifications (20x and 10x) on ROI detection accuracy.
• Demonstrated the method's utility in automated HER2 grading for breast cancer patients.

Main Results:

• Achieved high accuracy in ROI detection: 99% at 20x magnification and 97% at 10x magnification.
• The automated method significantly improved diagnostic agreement to 99.3% compared to clinical scores.
• Reduced the time for automated HER2 grading to just 15 seconds per case.

Conclusions:

• Automated ROI detection using Vision Transformer is a robust and accurate approach for WSI analysis.
• This method enhances consistency and efficiency in HER2 grading, addressing limitations of manual selection.
• The proposed technique holds significant potential for improving diagnostic workflows in digital pathology.