Hybrid cGAN-CNN Framework for Multi-Class Lung Cancer Detection Using SHAP, Grad-CAM, and t-SNE

Abstract

Lung cancer is one of the leading causes of cancer-related mortality worldwide, accounting for nearly 1.8 million deaths annually. Early and accurate diagnosis is essential for improving patient survival; however, the availability of annotated medical imaging data is limited and often suffers from class imbalance. To address these challenges, this study proposes an integrated framework combining a Conditional Generative Adversarial Network (cGAN), a Convolutional Neural Network (CNN), and Explainable Artificial Intelligence (XAI) techniques for automated lung cancer detection from Computed Tomography (CT) images. The dataset consisted of 1,097 CT images categorized into Normal, Benign, and Malignant classes, with an imbalanced distribution of approximately 400 Normal, 100 Benign, and 500 Malignant samples. Images were preprocessed by grayscale conversion, normalization, and resizing to 128 × 128 pixels. A cGAN was trained for 30 epochs to generate 200 synthetic CT images for data augmentation, while the CNN classifier was trained for 20 epochs for three-class classification. Experimental results demonstrated a classification accuracy of 68.64%, with Precision, Recall, and F1-score values of 0.959, 0.969, and 0.964 for the Normal class, respectively. The Benign class achieved an F1-score of 0.824, while the Malignant class obtained a Recall of 0.788. CNN training loss decreased from 1.0874 to 0.7706, indicating effective convergence. Explainability analysis using Grad-CAM, SHAP, and t-SNE confirmed that the model focused on clinically relevant regions such as pulmonary nodules and hilar structures. Although the generated images lacked complete anatomical realism, the proposed framework successfully demonstrated the feasibility of GAN-assisted lung cancer classification and established a strong baseline for future clinically deployable diagnostic systems.