Unsupervised single-image super-resolution for infant brain MRI

• Published in: NeuroImage (Orlando, Fla.) Vol. 317; p. 121293
• Main Authors: Tsai, Cheng Che, Chen, Xiaoyang, Ahmad, Sahar, Yap, Pew-Thian
• Format: Journal Article
• Language: English
• Published: United States Elsevier Inc 15.08.2025; Elsevier Limited; Elsevier
• Subjects: Babies; Brain; Brain - diagnostic imaging; Female; Humans; Image Processing, Computer-Assisted - methods; Infant; Infant brain MRI; Infant, Newborn; Infants; Magnetic resonance imaging; Magnetic Resonance Imaging - methods; Male; Neuroimaging; Neuroimaging - methods; Pediatrics; Radiology/Diagnostic Imaging; Random variables; Trustworthy reconstruction; Unsupervised Machine Learning; Unsupervised single-image super-resolution; Unsupervised single-image super-resolution; Trustworthy reconstruction; Infant brain MRI
• ISSN: 1053-8119; 1095-9572; 1095-9572
• DOI: 10.1016/j.neuroimage.2025.121293

Acquiring high-resolution (HR) MR images of infant brains is challenging due to lengthy scan times and limited subject compliance. Image super-resolution (SR) techniques can generate HR images from low-resolution (LR) inputs, reducing the need for extended acquisitions. However, most existing SR methods require HR images for training, limiting their practical use in real-world scenarios. To overcome this limitation, we propose an unsupervised single-image SR approach that requires only a single LR image for training. By integrating image space regularity with k-space consistency, our method enhances training stability and mitigates overfitting. Additionally, we introduce joint self-supervised learning to improve the fidelity of low-frequency content in the generated images. Our approach demonstrates both quantitative and qualitative improvements in MRI resolution for infants aged 1 week to 1 year, offering robust performance without manual hyperparameter tuning across diverse inputs. This innovation enables fully automated, high-throughput MRI resolution enhancement, addressing a critical need in pediatric neuroimaging. [Display omitted] •KC-DIP can be trained unsupervised on a single image for super-resolution.•KC-DIP employs various strategies to ensure good properties in the frequency space.•Training is stable with minimal hyperparameter adjustments across diverse images.•KC-DIP performs well on brain MR images of infants from 1 week to 1 year of age.