Hierarchical Deep Learning Framework with CBAM and TBAM techniques for Feature Fusion in Colon Cancer Classification from Histopathological Images

Colon or colorectal cancers remain a significant global health problem, with a growing number of cases and increasing mortality rates. This paper presents a hierarchical deep learning classification approach for the detection of colon cancer, employing top-down and bottom-up strategies. The framework integrates EfficientNet models with high-performance feature enhancement techniques, including the Convolutional Block Attention Module (CBAM) and Transformer-based attention mechanisms (TBAM), to improve feature extraction and classification accuracy. The model was trained and tested on a dataset of 25,000 histopathological images for improved accuracy. To analyze the efficiency of the proposed model, we used the performance metrics, including accuracy, G-Mean, F1-Score, Jaccard Index, Youden's Index and Diagnostic Odds Ratio (DOR). The experimental results indicate that the CBAM-based attention mechanism outperforms the Transformer-based attention model, with a classification accuracy of 87.42%, which proves CBAM's efficiency in enhancing deep-learning models for medical image analysis. The results of this paper indicate that the proposed deep learning model provides a robust and efficient solution for automatic colon cancer detection. By assisting pathologists in clinical decision-making, the model improves diagnostic accuracy, workload reduction, and timely and accurate cancer diagnosis.