Multi-class texture analysis in colorectal cancer histology

Area of Science:

• Digital Pathology
• Computational Pathology
• Medical Image Analysis

Background:

• Automatic tissue recognition is crucial for digital pathology.
• Current methods primarily focus on binary tumor/stroma separation, neglecting multi-class complexity.
• No prior work exists on multi-class texture separation for colorectal cancer histology.

Purpose of the Study:

• To develop and evaluate a multi-class classification strategy for histological tissue types in colorectal cancer.
• To introduce a novel dataset of 5,000 colorectal cancer histology images with eight distinct tissue types.
• To establish a new benchmark for multi-class tissue separation in this domain.

Main Methods:

• Compilation of a new dataset comprising 5,000 histological images of human colorectal cancer.
• Assessment of various texture descriptors and classifiers for multi-class classification.
• Development of an optimal classification strategy based on performance evaluation.

Main Results:

• The proposed optimal classification strategy significantly outperformed traditional methods.
• Tumor/stroma separation accuracy improved from 96.9% to 98.6%.
• A new state-of-the-art accuracy of 87.4% was achieved for eight-class multi-class tissue separation.

Conclusions:

• The developed method sets a new standard for multi-class tissue classification in colorectal cancer histology.
• The publicly available dataset will facilitate further research and benchmarking in digital pathology.
• This work advances the capability of automated analysis in complex histological samples.