Guided Diffusion: Balancing Fidelity and Diversity in Dataset Augmentation with EfficientNet and Gaussian Noise

The denoising diffusion probabilistic model (DDPM) has recently attracted massive attention due to its better capability of synthesizing high-quality and diverse synthetic data than generative adversarial network (GAN), paving the way for its application in data augmentation scenarios. However, balancing fidelity and diversity remains a challenge. To address the problem, a novel architecture is proposed, incorporating EfficientNet to extract features from the original dataset and fuse them with those of noise samples, guiding the denoising process and ensuring fidelity between synthetic samples and the original data. Additionally, random Gaussian noise is introduced to the UNet bottleneck output at each timestep to enhance diversity. A pre-trained CNN classification network follows to ensure label consistency between the reference and the synthetic images. The approach is evaluated through experiments on lung cancer prediction using a chest CT-scan dataset, achieving a 13.6% improvement in classification accuracy over baseline methods, 9.8% over the traditional cropping and rotation approach, and 4.1% over the GAN-based approach. These results validate the effectiveness of the proposed method for dataset augmentation.