PulmoNet: A Hybrid CNN-Vision Transformer Architecture for Enhanced Lung Nodule Classification in CT Imaging

Lung cancer remains the leading cause of cancer-related mortality, with early detection by CT screening being critical for patient survival. Current deep learning approaches face significant limitations: Convolutional Neural Networks (CNNs) extract local texture patterns but cannot capture global spatial relationships, while Vision Transformers (ViTs) model long-range dependencies but struggle with fine-grained feature extraction. Existing hybrid architectures use static fusion strategies that fail to adapt to diverse nodule characteristics. This paper presents PulmoNet, a novel hybrid framework that integrates a modified ResNet-50 with Vision Transformers through an adaptive cross-attention fusion mechanism that dynamically adjusts branch contributions based on individual nodule morphology. The framework processes CT volumes through parallel pipelines where CNNs extract multi-scale local patterns and transformers capture long-range spatial dependencies. Evaluated on the LUNA16 and LIDC-IDRI datasets using 5-fold cross-validation, PulmoNet achieves 94.7% accuracy and 0.982 AUC-ROC, outperforming state-of-the-art baselines by 3.5-5.4%. Cross-dataset evaluation demonstrates robust generalization across nodule sizes, types, and locations. PulmoNet demonstrates clinical viability with 93.8% sensitivity at 95% specificity and 143 ms inference time, establishing potential for real-time lung cancer screening programs.