An Integrated Deep Learning Framework Leveraging NASNet and Vision Transformer with MixProcessing for Accurate and Precise Diagnosis of Lung Diseases.
Authors
Affiliations (6)
Affiliations (6)
- Department of Information and Technology, Washington University of Science and Technology, Alexandria, VA 22314, USA.. Electronic address: [email protected].
- Department of Computer Software Engineering, Military College of Signals, National University of Sciences and Technology, Islamabad, Pakistan.. Electronic address: [email protected].
- Department of Computer Science, COMSATS University Islamabad, Vehari Campus, Pakistan. Electronic address: [email protected].
- Department of Computer Science, Kansas State University, Manhattan, KS 66506, USA. Electronic address: [email protected].
- School of Engineering and Technology Centre for Intelligent Systems, Central Queensland University, Australia. Electronic address: [email protected].
- SComputer Science Department, College of Science and Humanities, Imam Abdulrahman Bin Faisal University, Jubail 31961, Saudi Arabia. Electronic address: [email protected].
Abstract
Lung diseases such as pneumonia, tuberculosis, COVID-19, and lung cancer remain significant global health challenges that demand rapid and accurate diagnosis to improve patient outcomes. This study proposes NASNet-ViT, a novel deep learning framework that integrates the powerful convolutional feature extraction of NASNet with the global attention mechanisms of the Vision Transformer (ViT). To enhance diagnostic precision, a multi-stage preprocessing pipeline, termed MixProcessing, is introduced, combining wavelet transform decomposition, adaptive histogram equalization, and morphological filtering to improve image quality and feature clarity. The proposed NASNet-ViT model classifies lung images into five categories, normal, lung cancer, COVID-19, pneumonia, and tuberculosis achieving outstanding performance metrics: 98.9% accuracy, 0.99 sensitivity, 0.989 F1-score, and 0.987 specificity. Compared to established architectures such as MixNet-LD, D-ResNet, MobileNet, and ResNet50, NASNet-ViT demonstrates superior accuracy while maintaining a lightweight model size of only 25.6 MB and fast inference time of 12.4 seconds, making it practical for deployment in real-time, resource-constrained clinical environments. This research advances the field of medical image analysis by offering a robust and scalable AI solution capable of supporting clinicians in timely and precise lung disease diagnosis.