Sampling Based Tumor Recognition in Whole-Slide Histology Image With Deep Learning Approaches

Area of Science:

• Computational pathology
• Digital pathology
• Oncology research

Background:

• Histopathological identification of tumor tissue is crucial for cancer diagnosis.
• Whole-slide image (WSI) analysis presents challenges in efficient and spatial-correlated patch processing.
• Deep learning applications show promise in computational pathology.

Purpose of the Study:

• To develop a high-throughput system for precise tumor region detection in colorectal cancer histology slides.
• To improve the efficiency and accuracy of WSI analysis in cancer diagnostics.

Main Methods:

• Training a deep convolutional neural network (CNN) model.
• Employing a Monte Carlo (MC) adaptive sampling method for representative patch estimation.
• Integrating two conditional random field (CRF) models (correction and prediction) to capture spatial dependencies.

Main Results:

• The system achieved precise tumor region detection in colorectal cancer histology slides.
• Diagnostic time was reduced by 56.7% to 71.7% across slides with varying tumor distributions.
• Classification accuracy was increased by the proposed system.

Conclusions:

• The developed high-throughput system enhances the efficiency and accuracy of tumor detection in colorectal cancer WSI analysis.
• The integration of CNN, MC sampling, and CRF models offers a robust approach for computational pathology.
• This system has the potential to significantly aid pathologists in cancer diagnosis.