Magnetic resonance imaging (MRI) of the pancreas is one of the most advanced methods for evaluation of pancreatic parenchyma as well as pancreatic ducts [1]. Despite a rapid development in MR image quality in recent years, there are still challenges due to artifacts caused by physiologic motion in the upper abdomen, especially due to breathing [2]. These artifacts can deteriorate the image clarity and lead to reduction of anatomic detail [3, 4]. This is in particular relevant for dynamic contrast-enhanced (DCE) imaging that is of importance for further diagnostic work-up of pancreatic lesions [1, 5, 6]. Mostly a fast three-dimensional (3D) gradient echo (GRE) sequence (e.g., volumetric interpolated breath-hold examination [VIBE]) is used for non-enhanced as well as enhanced T1-weighted imaging including parallel imaging (PI). PI is the most common acceleration technique to perform imaging within one breath-hold to avoid the occurrence of breathing and motion artifacts. A significant disadvantage of PI is the loss of signal-to-noise ratio (SNR) as well as specific artifacts, which occur in particular at high acceleration factors (> 3–4), limiting the amount of acquisition time (TA) reduction.

A method called iterative denoising has been successfully applied to compensate for acceleration-induced SNR loss [7], [8], [9]. Another technique that has been gaining importance during the last years is called super-resolution showing promising results, especially in combination with deep learning approaches (e.g., in musculoskeletal, brain, and abdominal imaging) [10], [11], [12], [13], [14]. Super-resolution is a method capable of creating images with increased sharpness, SNR, and reduction of blurring and Gibb´s ringing using raw data of conventional acquisitions without the need of change of acquisition parameters leading to a potential improvement of image quality [15, 16]. This approach can also be used in addition to denoising. Furthermore, the combination of super-resolution post-processing including partial Fourier reconstruction settings allows a retrospective simulation of datasets with shorter acquisition time via simulation of more “aggressive” partial Fourier settings (e.g., 6/8 vs. 7/8).

The purpose of this study was to investigate the impact of deep learning-based super-resolution in combination with iterative denoising of T1-weighted VIBE Dixon imaging including simulated TA reduction via more aggressive partial Fourier factors on non-contrast and contrast-enhanced dynamic MRI of the pancreas regarding diagnostic confidence, lesion detectability, and image quality parameters.