Synthetic Microscopic Platelets Images Generation using WGAN-GP

Introduction: Data scarcity presents a major challenge in medical imaging. Generative Adversarial Networks (GANs) offer a promising solution by generating synthetic data to augment small datasets. Aim: The primary aim of this study is to evaluate the effectiveness of GAN-based synthetic data generation in enhancing CNN performance for platelet classification. Methods: The initial dataset contained 71 images, categorized into Control, Milrinone, and Zinc-plus-Milrinone classes. Using WGAN-GP, a synthetic dataset of 300 images was generated. These synthetic images were incorporated into the training of CNN models, including DenseNet121 and InceptionV3. Model performance was assessed comparing results from the original and augmented datasets. Results: In the non-augmented dataset, DenseNet121 achieved an accuracy of 81%, with 84% precision, 78% recall, and an F1-score of 81%. For InceptionV3, the non-augmented dataset yielded 82% accuracy, 80% precision, 76% recall, and an F1-score of 78%. However, after incorporating GAN-generated synthetic images, DenseNet121 achieved 97% accuracy, 97% precision, 95% recall, and a 96% F1-score, while InceptionV3 reached 94% accuracy, 93% precision, 90% recall, and 92% F1-score on the GAN-augmented dataset. Discussion/Conclusion: These results highlight the potential of GAN-generated synthetic data to significantly enhance the performance of CNN models in medical image classification, particularly in addressing the limitations of small datasets.