A Modified DCGAN for Breast Ultrasound Image Synthesis Based on Small Datasets

Generative Adversarial Networks (GANs) encounter several challenges such as the generation of unrealistic images, the production of synthetic images with limited diversity, and the risk of overfitting to the training dataset. These challenges hinder the extensive application of GANs in medical fields, where obtaining large medical image datasets is constrained by high costs, time limitations, and patient privacy concerns. To address these issues, we propose a modified Deep Convolutional GAN (DCGAN) model specifically designed for breast ultrasound (BUS) images under limited training data conditions. Our approach improves the consistency of the adversarial process and enhances the feature-capturing ability of GANs. The proposed model incorporates three key modifications and two promising training techniques: (1) replacing the rectified linear unit (ReLU) activation in the generator with the Scaled Exponential Linear Unit (SELU) activation, (2) implementing Spectral Normalization (SN), and (3) incorporating a Squeeze-and-Excitation (SE) block in the discriminator, (4) using label smoothing, and (5) applying asymmetric learning rate techniques during training. Evaluation results, based on Inception Score (IS), Mean Squared Error (MSE) metrics, and visual comparison, demonstrate that the proposed method outperforms traditional DCGAN architectures when trained on a limited dataset. These findings confirm that our approach provides a promising solution to the challenges of medical image synthesis in data-limited environments.