Transfer Learning Fusion Approaches for Colorectal Cancer Histopathological Image Analysis.

Area of Science:

• Medical Imaging
• Computational Pathology
• Artificial Intelligence in Medicine

Background:

• Accurate histopathological image classification is crucial for diagnosing colorectal cancer (CRC).
• Existing methods may benefit from enhanced feature extraction and focused analysis of critical image regions.

Purpose of the Study:

• To develop and evaluate novel attention-based decision fusion models for improved CRC histopathological image classification.
• To enhance feature extraction and model interpretability using spatial attention mechanisms.

Main Methods:

• Three attention-based decision fusion models were created by combining pre-trained Convolutional Neural Networks (CNNs) (Inception V3, Xception, MobileNet) with a spatial attention mechanism.
• An attention-driven fusion strategy weighted model predictions by relevance and confidence.
• Models were validated on a combined dataset of 17,531 CRC histopathological images from Oman and a public repository.

Main Results:

• The proposed models achieved high classification accuracy (98-100%), strong Matthews Correlation Coefficient (MCC) and Kappa scores, and low misclassification rates.
• Attention-based models demonstrated superior performance compared to individual transfer learning approaches (p = 0.009).
• Gradient-weighted class activation mapping visually confirmed the focus on diagnostically relevant image regions, enhancing model interpretability.

Conclusions:

• The developed attention-based decision fusion models show significant potential for accurate and robust classification of colorectal cancer histopathological images.
• These models offer enhanced diagnostic support and interpretability, valuable for research and clinical applications.
• The findings underscore the effectiveness of attention mechanisms and decision-level fusion in deep learning for medical image analysis.