NANO-HYBRID: A LIGHTWEIGHT INCEPTIONNEXT-ATTENTION NETWORK FOR EFFICIENT LUNG CANCER DIAGNOSTICS

Abstract

Lung cancer remains a leading cause of cancer-related mortality, necessitating diagnostic tools that are both accurate and computationally efficient for widespread deployment. While recent hybrid Deep Learning models have achieved high classification performance, they typically rely on heavy architectures (>18 million parameters) and extensive pre-training, limiting their applicability on resource-constrained edge devices. This study proposes a Nano-Hybrid architecture that integrates lightweight InceptionNeXt convolutions with global attention mechanisms, designed to be trained entirely from scratch. We evaluated the model on two diverse datasets: the IQ-OTH/NCCD (3-class) and a multi-class Chest CT dataset (4-class). Despite containing ~89% fewer parameters (2.03M) than comparable state-of-the-art baseline models, our approach achieved 95.41% accuracy on the IQ dataset, demonstrating that massive capacity is not strictly required for high-performance diagnostics. On the challenging multi-class Chest CT dataset, the model achieved 86.11% accuracy, with a notable 1.00 AUC (Area Under Curve) for normal cases, ensuring zero false positives in healthy screenings. Explainability analysis using Grad-CAM further validates that the model correctly prioritizes pulmonary nodule structures over background artifacts.