Abstract
<title>Abstract</title> <p>Authenticity in medical imaging is essential for accurate diagnosis and patient safety, yet recent advances in generative models have introduced risks of deepfake manipulation. This paper presents an enhanced EfficientNetV2-L framework for detecting deepfake lung CT images with improved accuracy and robustness. To strengthen the detection model, a balanced dataset of authentic and synthetic CT images is created, where the synthetic samples are generated using Stable Diffusion model, solely for training and evaluation purposes. Experiments on the IQ‑OTH/NCCD dataset, comprising 1,190 CT images across normal, benign, and malignant categories, confirm the robustness of the detection model. The proposed model is evaluated using training loss, precision, recall, F1-score, and mean Average Precision (mAP). Comparative analysis with existing deep learning models demonstrates faster convergence and superior detection performance.</p>