Related Experiment Video
Updated: Aug 24, 2025
15:17
Colorectal Cancer Cell Surface Protein Profiling Using an Antibody Microarray and Fluorescence Multiplexing
Published on: September 25, 2011
14.1K
Attention-guided sampling for colorectal cancer analysis with digital pathology
Andrew Broad1,2, Alexander I Wright3,4, Marc de Kamps1,2,5
1School of Computing, University of Leeds, Sir William Henry Bragg Building, Woodhouse Lane, Leeds LS2 9BW, UK.
Journal of Pathology Informatics
|October 21, 2022
Summary
This study introduces an attention-inspired algorithm for pathology image analysis, significantly speeding up whole-slide image processing and improving tumor region identification. The new method enhances efficiency and accuracy in computational pathology research.
Area of Science:
- Computational pathology
- Digital pathology
- Medical image analysis
Background:
- Convolutional neural networks (CNNs) face limitations processing whole-slide images (WSIs) due to small patch sizes, making tile-by-tile analysis slow and inefficient.
- Current methods are susceptible to noise from uninformative image areas, impacting diagnostic accuracy in digital pathology.
Purpose of the Study:
- To develop a novel attention-inspired algorithm for efficient and accurate pathology image analysis.
- To improve the selection of informative image patches from WSIs for CNN processing.
- To enhance the speed and accuracy of tumor detection and analysis in digital pathology.
Main Methods:
- An attention-inspired algorithm was developed to intelligently select informative image patches from WSIs using sparse, randomized grid patterns and iterative resampling.
- A two-stage CNN model, adapted from brain imaging, was employed for classifying tumor regions and distinguishing false-positive normal epithelium.
- The pipeline integrated patch classification, region of interest (ROI) prediction, and tumor-stroma ratio (TSR) evaluation.
Main Results:
- The attention-based sampling pipeline achieved a mean F1 (Dice) score of 86.6% for ROI prediction across 100 WSIs.
- Tumor-stroma ratio (TSR) calculation using the pipeline showed a lower root mean square (RMS) error (11.3%) compared to a standard tile-by-tile approach (13.5%).
- The proposed pipeline processed WSIs 3.3x to 6.3x faster than conventional tile-by-tile methods.
Conclusions:
- The attention-based sampling pipeline offers a significant advancement in computational pathology, improving efficiency and accuracy in WSI analysis.
- This method provides a valuable tool for pathology researchers, enabling faster and more precise identification of tumor regions and facilitating further diagnostic calculations.
- The algorithm demonstrates potential for integration into broader digital pathology workflows for enhanced patient care and research.
