Abstract:
Lung cancer continues to be a predominant source of cancer-related mortality, with early identification essential for enhancing patient outcomes. This paper offers a comparative assessment of deep learning architectures, including VGG16, ResNet50, MobileNetV2, and Vision Transformer (ViT), for the identification of lung cancer from CT scans. To improve diagnostic precision, we propose an attention-enhanced VGG16 (AttVGG16) that incorporates the Convolutional Block Attention Module (CBAM) to emphasise prominent problematic areas. Experimental results indicate that baseline VGG16 outperforms conventional architectures, whereas ResNet50, MobileNetV2, and ViT display diminished predictive efficacy due to constrained feature representation or data prerequisites. AttVGG16 surpasses all baseline models, with 97.78% accuracy, 97.80% sensitivity, 97.76% F1 score, and 0.9178 MCC, underscoring the effectiveness of attention processes in accentuating diagnostically pertinent areas and minimising false negatives. The research employed a meticulously curated dataset of lung CT scans, incorporating extensive preprocessing such as normalisation, augmentation, and class balancing to mitigate data scarcity and improve model generalisability. Performance was assessed through many measures to ensure a comprehensive evaluation of categorization accuracy, sensitivity, and predictive reliability. Moreover, the attention mechanism integrated into AttVGG16 enables the model to emphasise significant areas in CT images, enhancing the network's interpretability for clinical applications. These findings highlight the efficacy of attention-enhanced CNNs for accurate and early lung cancer identification in CT imaging, providing a valuable resource to assist radiologists in diagnostic decision-making. The suggested methodology may aid in diminishing diagnostic inaccuracies, enabling prompt interventions, and eventually enhancing patient management and outcomes. Future endeavours may encompass the expansion of this framework to encompass larger and more heterogeneous datasets, the integration of multimodal imaging, and the creation of real-time clinical decision support systems for the automated identification of lung cancer.
