DeepCRC-Net: An Attention-Driven Deep Learning Network for Colorectal Cancer Classification Using Xception and Efficient Lightweight Local Feature Fusion Networks

Abstract

Colorectal cancer (CRC) is one of the primary causes of cancer-related deaths globally, and early detection is critical for improving patient outcomes. Histopathological analysis of colorectal tissue remains the gold standard for diagnosis, however it is labour-intensive and prone to human error which could lead to misdiagnosis. Deep learning models offer a promising solution by automating the analysis and enhancing diagnostic accuracy. This research introduces DeepCRC-Net, a novel dual-track architecture designed to classify CRC histopathology images using the EBHI dataset. The first track leverages the Xception network to capture long-range dependencies using depth-wise separable convolutions and residual connections. The second track employs the Efficient Lightweight Local Feature-Fusion Network (ELLFFN), which integrates Efficient Semi-Local Attention Convolution (ESAC) and Dynamic Deformed Shuffle-Fusion Convolution (DDSFC) blocks to capture semi-local and local features efficiently. The extracted features from both tracks are concatenated and refined using Shuffle Attention, which enhances the ability of the model to focus on the most informative regions of the CRC images. The proposed model achieved an accuracy of 98.6% and an F1-score of 99.25% when tested on the EBHI dataset. DeepCRC-Net outperforms existing studies by 2.5% in accuracy and state-of-the-art CNNs by 4.2%.