Abstract
Lung cancer continues to be a predominant cause of cancer-related mortality globally. In 2022, lung cancer accounted for around 2.5 million new cases and around 1.8 million fatalities, highlighting the necessity for precise and effective computer-aided diagnostics. Timely detection is especially vital for non-small cell lung cancer (NSCLC), which constitutes roughly 80–85% of all lung cancer instances, but continues to pose difficulties in standard clinical practice. This paper presents a Double Attention Hybrid CNN-HiFuse architecture for categorizing lung cancer into three classes (normal, benign, malignant) using chest Computed Tomography (CT) images. The model is trained and evaluated on the publicly accessible IQ-OTH/NCCD dataset, which comprises 1,190 CT slices from 110 patients, using a defined preprocessing and augmentation protocol to address class imbalance. The proposed Hybrid CNN-HiFuse, which combines multi-scale feature fusion with channel and spatial attention methods, is evaluated against a bespoke CNN and transfer-learning benchmarks (VGG16, ResNet50). The model attains an overall classification accuracy of 98.12% on the stated test split, with precision, recall, and F1-score of 98.17%, 98.12%, and 98.13%, respectively, surpassing the baseline designs. The confusion matrix and ROC studies demonstrate minimal misclassification rates, especially for malignant nodules, while the attention maps emphasize clinically significant areas, hence improving interpretability. The present assessment, confined to a modest single-centre CT dataset, indicates that the Double Attention Hybrid CNN-HiFuse framework is a viable candidate for incorporation into clinical decision-support systems, potentially enhancing radiologists’ efficacy in early lung cancer detection. SUPPLEMENTARY INFORMATION: The online version contains supplementary material available at 10.1038/s41598-026-42290-9.