A comprehensive review of attention-based multimodal deep learning techniques for lung cancer diagnosis

Lung cancer remains the deadliest cancer type worldwide, presenting a wide spectrum of subtypes, including several rare and difficult-to-diagnose types. Improving diagnostic accuracy and efficiency with technology seems promising, which helps medical experts to save many patients’ lives. Deep learning has emerged as a powerful tool for enhancing several critical lung cancer diagnostic tasks, including early detection, subtype classification, survival prediction, and tumor segmentation. Most recent developments are multimodal deep learning (MMDL) techniques, which aggregate information from multiple data types, including medical imaging, genomic data, and clinical records, achieving superior performance over single-modality approaches. Fusion strategies are developed to integrate features from multiple modalities effectively, while attention mechanisms further enhance performance by identifying diagnostically relevant features and dynamically weighting modalities. The reviewed studies suggest that attention-based MMDL can enhance performance across several lung cancer diagnostic and prognostic tasks by leveraging complementary information from heterogeneous data sources. While some studies reported substantial gains, particularly in treatment response and survival prediction, direct quantitative comparison remains challenging because of variations in datasets, clinical objectives, baseline models, and validation protocols. This paper presents a comprehensive review of attention-based MMDL approaches for lung cancer diagnosis, analyzing studies published between 2020 and 2025, with a focus on clinical applications, data modalities, fusion strategies, and attention mechanisms.