Enhanced DenseNet121-Based Framework with MPCNN-TAO and SE Modules for Early Gastric Cancer Detection in Histopathological Images

Abstract

Accurate and early detection of gastric cancer from histopathological images remains a significant challenge due to complex tissue structures and subtle morphological variations. This study proposes an advanced deep learning architecture by enhancing the DenseNet121 backbone with a custom Hyper Model that integrates Multi-Path Convolutional Neural Network with Transformer-Attention Optimization (MPCNN-TAO), Multi-Path Feature Extraction, and Squeeze-and-Excitation (SE) layers. The MPCNN-TAO module enables the model to capture global contextual dependencies while preserving essential spatial information through multi-head self-attention and convolutional fusion. The Multi-Path Feature Extraction block aggregates fine and coarse features using parallel convolutions of varying kernel sizes, enabling the network to better learn discriminative patterns across heterogeneous tissue regions. Additionally, SE layers are incorporated to adaptively recalibrate channel-wise features, improving the network’s focus on salient regions associated with malignancy. Experimental results on benchmark gastric cancer histopathology datasets demonstrate that the proposed model outperforms standard CNN architectures, achieving superior classification accuracy and interpretability. The hybrid framework provides a robust and scalable solution for aiding pathologists in the early diagnosis of gastric cancer.