Hybrid deep learning for interpretable lung cancer recognition across computed tomography and histopathological imaging modalities

Lung cancer is one of the deadliest cancers globally, primarily due to its silent development and difficulties with late diagnoses. Traditional diagnostic methods like manual CT scans and histopathological slide analysis face challenges such as observer variability, limited sensitivity, and difficulties in handling large volumes of data. Convolutional neural networks (CNNs) can automate image classification, but their limited receptive fields hinder complex tissue structure analysis. Vision Transformers (ViTs) provide a solution but typically need large datasets and significant computing power, making them impractical in clinical settings. Our study introduces a hybrid deep learning (DL) framework using the LEViT architecture to enhance lung cancer classification. We utilized two public datasets: IQ-OTH/NCCD, containing 1,097 CT images categorized into normal, benign, and malignant, and another dataset with 25,000 histopathological images across five tissue types. Our methodology included a multi-stage preprocessing pipeline to resize, reduce noise, enhance contrast, normalize, and augment data to tackle class imbalance and improve generalization. We evaluated our model using metrics like accuracy, F1 score, specificity, PR AUC, and Matthews Correlation Coefficient (MCC) through 10-fold stratified cross-validation. Our LEViT-based model surpassed top models such as CoAtNet and CrossViT, achieving 99.43% accuracy and 98.36% MCC on the IQ-OTH/NCCD dataset, and 99.02% accuracy with 97.97% MCC on the other dataset. Additionally, we developed a real-time web application for clinicians to upload images and receive visual explanations via Grad-CAM, promoting transparency in decision-making. This work provides a scalable, accurate, and explainable AI solution for lung cancer recognition, connecting high-performance algorithms with clinical practice.