Classification of colorectal tissue images from high throughput tissue microarrays by ensemble deep learning methods

Area of Science:

• Digital pathology
• Computational biology
• Artificial intelligence in medicine

Background:

• Tissue microarrays (TMAs) are valuable for AI but susceptible to sectioning errors requiring re-labeling.
• High-throughput TMAs present challenges for accurate tissue classification due to potential content shifts.

Purpose of the Study:

• To investigate ensemble deep learning methods for accurate colorectal tissue classification in high-throughput TMAs.
• To develop a robust algorithm for classifying tumor, normal, and "other" tissues from H&E stained TMA core images.

Main Methods:

• Extracted Hematoxylin and Eosin (H&E) core images from three international cohorts (n=15,150 cores).
• Applied five independent and ensemble deep learning models after color enhancement.
• Utilized ground-truth data (n=8689 cores) and test augmentation for validation.

Main Results:

• The Soft Voting Ensemble of VGG and CapsNet models achieved the highest predictive accuracy: 0.982 (normal), 0.947 ("other"), and 0.939 (tumor).
• This ensemble method outperformed independent and single base-estimator ensemble approaches.
• The algorithm demonstrated robustness across different institutions, core sizes, and staining intensities.

Conclusions:

• A high-accuracy AI algorithm for colorectal tissue classification in high-throughput TMAs was developed.
• The developed method reduces errors in TMA core evaluations, even with pre-existing labels.
• This approach is applicable to diverse TMA datasets, enhancing diagnostic reliability.