Attn-DeCGAN: A Diversity-Enhanced CycleGAN With Attention for High-Fidelity Medical Image Translation

Abstract

Unpaired image-to-image translation has emerged as a transformative paradigm in medical imaging, enabling unpaired image translation without the need for aligned datasets. While cycle-consistent generative adversarial networks (CycleGANs) have shown considerable promise in this domain, they remain inherently constrained by the locality of convolutional operations, resulting in global structural inconsistencies, and by mode collapse, which restricts generative diversity. To overcome these limitations, we propose Attn-DeCGAN, a novel attention-augmented, diversity-aware CycleGAN framework designed to enhance both structural fidelity and perceptual diversity in CT-MRI translation tasks. Attn-DeCGAN replaces conventional ResNet-based generators with Hybrid Perception Blocks (HPBs), which synergise depthwise convolutions for spatially efficient local feature extraction with a Dual-Pruned Self-Attention (DPSA) mechanism that enables sparse, content-adaptive modeling of long-range dependencies at linear complexity. This architectural innovation facilitates the modeling of anatomically distant relationships while maintaining inference efficiency. The model is trained using a composite loss function incorporating adversarial, cycle-consistency, identity, and VGG19-based structural consistency losses to preserve both realism and anatomical detail. Extensive empirical evaluations demonstrate that Attn-DeCGAN achieves superior performance across key metrics, including the lowest FID scores (60, 58), highest PSNR (27, 33), and statistically significant improvements in perceptual diversity (LPIPS, ) compared to state-of-the-art baselines. Ablation studies underscore the critical role of spectral normalization in stabilizing adversarial training and enhancing attention effectiveness. Expert radiologist assessments confirmed the clinical superiority of Attn-DeCGAN over the next best baseline, DeCGAN, with 100% real classifications and higher confidence scores in CT synthesis, and more anatomically convincing outputs in MRI translation. This has particular utility in low-resource clinical environments where MRI is scarce, supporting synthetic MRI generation for diagnosis, radiotherapy planning, and medical image dataset augmentation. Despite increased training complexity, Attn-DeCGAN retains efficient inference, positioning it as a technically robust and clinically deployable solution for high-fidelity unpaired medical image translation.