PABCNet: A super-resolution network for non-contrast coronary MRA with inter-slice fusion and Bayesian confidence...

Non-contrast coronary magnetic resonance angiography (NC-CMRA) suffers from inherent low spatial resolution, whereas existing deep learning-based super-resolution methods offer no reliability assessment, which hinders their clinical adoption.

To bridge this gap, we develop a task-tailored integrated super-resolution pipeline named PABCNet, equipped with inter-slice phase alignment and calibrated Bayesian confidence quantification. PABCNet ensures data fidelity through a physics-driven low-resolution image synthesis pipeline that simulates k-space undersampling and complex noise. Crucially, to reconstruct anatomically continuous coronary structures, we configure an inter-slice feature propagation module with frequency-domain phase alignment and deformable convolution to achieve sub-pixel registration. Additionally, a Bayesian confidence quantification module employs Monte Carlo Dropout and Laplacian likelihood to jointly model epistemic and aleatoric uncertainty and generate pixel-wise confidence maps.

Extensive experiments on a private Coronary dataset and the public CardioScans dataset demonstrate that PABCNet achieves state-of-the-art performance, with a PSNR of 35.83 dB and an SSIM of 0.9729. Notably, it achieves a vessel-region ROI-PSNR of 32.68 dB (2.18 dB higher than the second-best method). Its reliability assessment yields a low expected calibration error of 0.0234 and effectively distinguishes reliable from uncertain regions. Radiologist evaluation confirms its significant superiority in coronary clarity, cardiac structure depiction and artifact suppression over both comparative methods and original high-resolution scans.

Without increasing the scanning burden, the proposed method realizes high-quality super-resolution reconstruction and reliability assessment of NC-CMRA, and provides auxiliary information for the image quality assessment of non-invasive coronary artery disease.