LCDLNet: Attention Mechanism‐Based Deep Learning Framework for Lung Cancer Diagnosis from Computed Tomography (CT) Images

This paper introduces a new deep learning architecture, LCDLNet, that can diagnose and classify lung cancer using CT scan images. LCDLNet makes use of a custom CNN structure including residual connections, DenseNet121, SE blocks, and transformer modules, which allow multi‐scale and contextualized feature mappings. It combines the intermediate feature layers from CNN and DenseNet121 to provide a comprehensive solution to the extraction of distinguishing features necessary for the accurate classification of lung diagnosis. The LIDC‐IDRI dataset was used to train LCDLNet and evaluate its performance. The training dataset consisted of 5292 CT slices obtained from 165 patients, and all images were resized to 224 224 pixels and normalized to achieve stable training. The results achieved with LCDLNet were excellent, with a training accuracy of 99.53%, validation accuracy of 98.98%, and test accuracy of 99.81%. Achieving 99.86% Precision, 99.67% Recall, and 99.80% AUC scores demonstrates that this model has been able to generalize successfully while exhibiting minimal overfitting. The proposed framework is effective at solving class imbalance issues while providing a unique attention mechanism. This will provide an important step towards building an automated system for the diagnosis of lung cancer and the development of a clinically useful decision support tool.