A Dual-Path Deep Learning Approach for Robust Classification of Colon Histopathology Images Using Attention Mechanisms: A Multicenter Study

Background and Aims: Colorectal cancer is among the most prevalent and lethal malignancies globally. Histopathological image analysis is essential for accurate diagnosis, yet limited resources and inter-observer variability often challenge it. Existing deep learning models are typically trained on homogeneous public datasets that lack the staining, scanning, and demographic diversity encountered in everyday clinical settings, particularly in low-resource environments. These limitations reduce their generalizability and clinical relevance. In this study, we aimed to develop a robust and generalizable deep learning model for binary classification of H&E-stained colon tissue images into adenocarcinoma and normal categories. We combined ensemble learning with a diverse multicenter dataset to enhance real-world applicability. Methods: We curated a private dataset of over ۲,۰۰۰ H&E-stained colon tissue images from three independent pathology laboratories, capturing variation in staining protocols, microscope cameras, and patient demographics. Expert pathologists annotated all images. We separately trained three pre-trained convolutional neural networks—DenseNet۱۲۱, InceptionV۳, and VGG۱۹—on a publicly available dataset to capture generalizable features. We then fine-tuned each model on our multicenter dataset. To improve performance, we integrated the three networks using a Squeeze-and-Excitation block to enhance feature attention. The merged outputs from each path were passed through fully connected layers for final binary classification. Preprocessing steps included tile extraction, normalization, and data augmentation. We evaluated model performance using accuracy, precision, recall, and F۱-score. Results: The proposed model achieves superior performance compared to individual models (VGG۱۹, InceptionV۳, DenseNet۱۲۱). On the test set, it reached an accuracy of ۰.۹۹۱۲ and F۱-score of ۰.۹۹۲۱, outperforming VGG۱۹ (acc: ۰.۹۶۹۴, F۱:)