Improving performance in colorectal cancer histology decomposition using deep and ensemble machine learning.

Area of Science:

• Digital pathology
• Computational oncology
• Biomarker discovery

Background:

• Hematoxylin and eosin stained histologic samples are standard in colorectal cancer management.
• Current gold standards for patient stratification rely on costly and time-consuming genetic tests.
• Convolutional neural networks (CNNs) show promise for extracting biomarkers from readily available histologic images.

Purpose of the Study:

• To develop and evaluate a novel hybrid deep transfer learning and ensemble machine learning model for accurate tissue classification in whole slide images.
• To enhance the prognostic potential of imaging-based biomarkers for colorectal cancer patient stratification.
• To provide a faster, automated, and cost-effective alternative to traditional genetic biomarker discovery.

Main Methods:

• A hybrid model combining EfficientNetV2 architecture with a random forest classification head was developed.
• The model leverages deep transfer learning and ensemble machine learning techniques.
• Performance was evaluated on internal and external test sets, comparing against transformer and neural architecture search baselines.

Main Results:

• The hybrid model achieved 96.74% accuracy on the external test set (95% CI: 96.3%-97.1%).
• The model demonstrated exceptionally high accuracy of 99.89% on the internal test set.
• The developed model significantly improves upon previous approaches for tissue class decomposition.

Conclusions:

• CNN-based biomarkers extracted from histologic images can predict patient outcomes comparably to gold standards.
• Accurate tissue class decomposition is crucial for enhancing the prognostic power of imaging-based biomarkers.
• The publicly available model offers a promising, efficient, and cost-effective tool for colorectal cancer biomarker discovery and patient stratification.