A Novel Quad-Path GAN for Bidirectional Brain Image Translation Between CT and MRI

Unpaired cross-modality image translation is an important yet challenging task in medical image analysis. Most existing approaches are restricted to single-pair domain mappings and often struggle to maintain anatomical fidelity and perceptual consistency across modalities. These limitations can hinder the clinical applicability of synthetic images, particularly in settings that rely on multi-sequence or multi-domain information. In this work, we propose Quad-(Cycle)GAN , a multi-domain generative framework that enables unpaired image-to-image translation across four imaging domains simultaneously. Unlike conventional models that treat each domain pair separately, Quad-(Cycle)GAN introduces a unified architecture comprising four generators and four discriminators. The design incorporates redundant cyclic supervision, multi-scale adversarial learning, and domain-aware consistency constraints to support robust translation while preserving anatomical structure and inter-slice coherence. We train a Quad-(Cycle) GAN on unpaired CT and MRI datasets and evaluate its performance using a combination of perceptual metrics (FID, KID, LPIPS) and clinically relevant image quality measures, including brightness stability, contrast sharpness, and artifact suppression. Our results show that Quad-(Cycle)GAN outperforms standard Cycle-GAN models across all metrics. Notably, it improves brightness consistency by 53% (p < 0.01) and achieves a large perceptual effect size (Cohen’s d = 1.23). Visual assessments further confirm smoother anatomical transitions and clearer tissue boundaries. These findings suggest that Quad-(Cycle)GAN offers a scalable and clinically meaningful solution for multi-domain medical image synthesis. By producing perceptually coherent and anatomically accurate outputs across multiple modalities, our framework has the potential to support a range of downstream applications, from diagnostic assistance to image pre-processing in multimodal workflows.