Colon histology slide classification with deep-learning framework using individual and fused features.

Area of Science:

• Medical Imaging
• Computational Pathology
• Artificial Intelligence in Oncology

Background:

• Colorectal cancer (CRC) incidence is rising globally, posing a significant diagnostic challenge.
• Early detection through screening and timely treatment are crucial for managing CRC.
• Current diagnostic methods require efficient and accurate tools for histology slide analysis.

Purpose of the Study:

• To develop and evaluate a Deep-Learning Framework (DLF) for classifying colon histology slides into normal and cancer categories.
• To leverage deep-learning-based features for automated cancer detection.
• To assess the clinical significance of the proposed DLF.

Main Methods:

• Image collection, resizing, and pre-processing of 4000 histology slides (2000 normal, 2000 cancer).
• Extraction of Deep Features (DF) using selected schemes.
• Binary classification using a 5-fold cross-validation, comparing individual, fused, and ensemble DFs.
• Verification of results using binary classifiers, including K-Nearest Neighbor (KNN).

Main Results:

• Fused Deep Features (DF) achieved a high detection accuracy of 99% when using the K-Nearest Neighbor (KNN) classifier.
• Individual DFs yielded a classification accuracy of 93.25%.
• Ensemble DFs resulted in a classification accuracy of 97.25%.

Conclusions:

• The developed Deep-Learning Framework (DLF) demonstrates high efficacy in classifying colon histology slides.
• Fused Deep Features (DF) offer superior performance for accurate colorectal cancer detection.
• This approach holds potential for improving early cancer screening and reducing diagnostic workload.