Adaptive Elastic GAN for High-Fidelity Blood Cell Image Hallucination and Classification

Automated blood cell classification is crucial for hematological analysis, yet the scarcity of annotated medical datasets challenges deep learning models. This study presents a novel semi-supervised Elastic Generative Adversarial Network that enhances classification accuracy, improves synthetic image generation quality, and reduces computational complexity. The model utilizes image hallucination, which generates artificial but realistic images to supplement datasets, addressing data imbalance and improving model generalization. The generated images exhibit high fidelity, meaning they closely resemble real blood cell images in texture, morphology, and structural details, ensuring their effectiveness in training deep learning models. Compared to the conventional semi-supervised Wasserstein Generative Adversarial Network, the proposed model achieves a 5.5% increase in classification accuracy, a 34.5% reduction in computational complexity, and a 21.0% decrease in training time, demonstrating superior efficiency and scalability. In addition, it greatly enhances Fréchet Inception Distance and Wasserstein distance, which verifies improved realism and diversity of synthetic samples. Complementing pruning-based optimization ensures reduced model deployment size, rendering it appropriate for real-time medical diagnosis. Through image hallucination, the proposed method sufficiently alleviates data deficiency and improves the reliability of automated blood cell classification, qualifying it as a prospective tool for advanced hematological studies and diagnosis purposes.