Unpaired image-to-image translation, demonstrated by frameworks like CycleGAN, has shown promise in critical applications, including medical imaging, but is hindered by challenges such as mode collapse and structural degradation. To address these, we propose DeCGAN, a Diversity-enhanced CycleGAN framework that integrates spectral normalisation with a novel composite loss function. This loss function combines structural consistency loss, leveraging high-level feature representations from a pre-trained VGG19 network, with mode-diversity loss to enhance both structural details and output diversity. Extensive evaluations on unpaired CT and MRI datasets demonstrate DeCGAN’s superiority over state-of-the-art models. DeCGAN achieved Inception Scores (IS) of 2. 44 (CT-to-MRI) and 3. 94 (MRI-to-CT), marking a 75% improvement over CycleWGAN-GP. Structural fidelity metrics also showed SSIM scores of 0. 94 (CT-to-MRI) and 0. 98 (MRI-to-CT), and PSNR improvements of 11. 9% and 17. 4%, respectively. DeCGAN also exhibited robust diversity with higher LPIPS scores of 0. 473 (CT-to-MRI) and 0. 375 (MRI-to-CT), and consistently low FID scores (67 and 70). Statistical analysis confirms DeCGAN’s significant gains in structural fidelity and diversity (p < 0. 05). Qualitative evaluation by an expert radiologist confirmed the clinical relevance of the generated images, with 36. 7% of CT and 33. 3% of MRI images classified as real clinical images, and over 70% of these receiving high confidence scores (4-5 on a 5-point scale), demonstrating the presence of clinically relevant features. DeCGAN demonstrates consistent improvements over existing CycleGAN-based approaches, offering a balanced trade-off between diversity, structural fidelity, and clinical plausibility in unpaired CT-MRI translation.
Cobbinah et al. (Mon,) studied this question.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: