TransferGAN-Transformer Image Generation for Leukemia Microscopy Analysis

Abstract

Generative Adversarial Networks (GANs) have shown strong potential for medical image generation, but domain adaptation, feature transfer, and medically meaningful evaluation remain challenging in leukemia microscopy. This paper describes a TransferGAN-Transformer Hybrid architecture for AI-based microscopy image synthesis, a topic aligned with intelligent image processing and machine-learning systems. The architecture combines transfer learning with transformer attention to generate leukemia image samples for controlled augmentation and computational analysis. The study has four objectives: to design a transfer-attention generator, improve structural fidelity in generated leukemia images, compare adversarial architectures under common image-quality metrics, and introduce a Composite Medical Synthesis Metric (CMSM) that combines structural similarity, signal fidelity, distributional realism, and morphology relevance. The model is evaluated on a medical imaging dataset with 10,661 training images. It achieves a Fréchet Inception Distance (FID) of 50.08, Structural Similarity Index (SSIM) of 0.763, Peak Signal-to-Noise Ratio (PSNR) of 26.01 dB, Learned Perceptual Image Patch Similarity (LPIPS) of 0.056, and Kernel Inception Distance (KID) of 0.025 at optimal checkpoints. Results show that transfer learning and transformer attention improve structural fidelity and feature adaptation compared with the TransferGAN baseline, while CMSM provides a domain-aware objective for interpreting broader GAN-family comparisons in biomedical image-processing workflows.