Abstract Objectives: To develop an attention-guided GAN framework for brain MRI super-resolution that increases clinical interpretability and anatomical detail while improving diagnostic accuracy from low-resolution data. Methods: For brain MRI super-resolution, an Attention-based Generative Adversarial Network (AttentionGAN) that integrates channel and spatial attention was created. ADNI was used to train on multi-center datasets, while OASIS was used for testing and validation. Controlled downsampling was used to simulate low-resolution images. Performance was assessed using PSNR and SSIM, compared to super-resolution baselines based on CNN and GAN, and qualitatively validated by expert review and attention map analysis. Findings: The proposed AttentionGAN demonstrated a consistent and statistically significant improvement in brain MRI super-resolution performance across all evaluated datasets. When compared to traditional interpolation, CNN-based, and existing GAN-based techniques, AttentionGAN showed better structural fidelity and contrast preservation, with an average PSNR improvement of 2.1 to 2.8 dB and an SSIM increase of 0.03 to 0.06. Qualitative analysis addressed the over-smoothing and hallucination problems frequently reported in previous GAN-based approaches, revealing sharper tumor margins and clearer gray-white matter differentiation. The model consistently focused on clinically relevant anatomical regions, as confirmed by attention map visualization - a feature that is mostly lacking in current approaches. These results corroborate recent studies that emphasize the advantages of attention mechanisms in medical image analysis and expand on them by specifically matching attention to diagnostic regions. Overall, the findings contribute to the body of literature by showing that attention-guided super-resolution can simultaneously increase clinical interpretability and quantitative image quality, boosting diagnostic confidence in practical MRI applications. Novelty: In order to overcome the limitations of current approaches in terms of hallucination and diagnostic reliability, this study presents an attention-aligned GAN that integrates spatial and channel attention to produce clinically interpretable brain MRI super-resolution. Keywords: Brain MRI, Super-Resolution, AttentionGAN, Spatial Attention, Channel Attention
Tivari et al. (Sun,) studied this question.