Advanced artificial intelligence driven framework for lung cancer diagnosis leveraging SqueezeNet with machine learning algorithms using transfer learning

Lung cancer is a severe global public health problem, and early detection is important for improving patient wellbeing. This research presents an advanced Artificial Intelligence (AI) driven framework that integrates deep learning and machine learning techniques to enhance lung cancer classification in chest Computed Tomography (CT) scans. Leveraging transfer learning, SqueezeNet a lightweight Convolutional Neural Network (CNN) is employed for feature extraction, which is then processed by Machine Learning (ML) classifiers. A dataset comprising 950 chest scans from 110 test cases is used to classify tumors into benign, malignant, and normal categories. Among the tested models, SqueezeNet combined with Logistic Regression (LR) achieves the highest accuracy of 92.9%. Performance evaluation is conducted using multiple classification metrics, including Confusion Matrix and Calibration plots, demonstrating the model’s reliability in early lung cancer detection. The proposed AI-driven hybrid framework offers a promising approach to improving diagnostic accuracy, ultimately benefiting both patients and the healthcare system. • Hybrid AI Framework: To effectively diagnose lung cancer from CT images, a unique AI-driven framework was developed that combines SqueezeNet (for deep feature extraction) with machine learning classifiers (LR, SVM, RF) using transfer learning. • Good SVM Performance: SVM was the most dependable model for classification, achieving the best results on all measures, including the lowest Log Loss (0.134), F1-score (0.952), and AUC (0.989). • Comprehensive Evaluation: To confirm model correctness, calibration, and generalization ability, several evaluation tools were used, including confusion matrices, calibration plots, and important metrics (AUC, F1, and MCC). • Significant Improvement Over Current Models: The suggested model demonstrated a 10.84% improvement over AlexNet and a 58.62% F1-score improvement over solo SqueezeNet, demonstrating the effectiveness of the hybrid technique.