Abstract
INTRODUCTION: Oral squamous cell carcinoma (OSCC) is a significant global health burden, where timely and accurate diagnosis is essential for improved patient outcomes. Conventional diagnosis relies on manual evaluation of hematoxylin and eosin (H&E)-stained slides, a time-consuming process requiring specialized expertise and prone to variability. While deep learning methods, especially convolutional neural networks (CNNs), have advanced automated analysis of histopathological images for cancerous tissues in various body parts, OSCC presents unique challenges. Its infiltrative growth patterns and poorly defined boundaries, coupled with the complex architecture of the oral cavity, make accurate segmentation particularly difficult. Traditional CNNs which sturggle to capture critical global contextual information often fail to distinguish the complex tissue structures in OSCC images. METHODS: To address these challenges, we propose a novel architecture called gamUnet, which integrates the Global Attention Mechanism (GAM) to enhance the model's ability to capture global cross-modal information. This allows the model to focus on key diagnostic regions while retaining detailed spatial information. Additionally, we introduce an extended model, gamResNet, to further improve OSCC detection performance. Both architectures show significant improvements in handling the unique challenges of oral cancer images. RESULTS: Extensive experiments on public datasets show that our GAM-enhanced architecture significantly outperforms conventional models, achieving superior accuracy, robustness, and efficiency in OSCC diagnosis. DISCUSSION: Our approach provides an effective tool for clinicians in diagnosing OSCC, reducing diagnostic variability, and ultimately contributing to improved patient care and treatment planning.