In patients with cancer, diffusion-weighted magnetic resonance imaging (DWI) represents a useful tool to visualize the main tumor and to evaluate metastases, as well as to monitor tumor progression and response to treatment, by exploiting the property of restricted water diffusion in tumors [1]. However, utilization of DWI in the abdomen and pelvis is challenged by reduced signal-to-noise ratio (SNR) at high diffusion weighting (or b-value) and increased geometric distortions caused by magnetic field inhomogeneities [2]. A common solution to increase SNR is to repeat the high b-value acquisition several times and average the different repetitions, which results in extended scan times. Multi-shot echo-planar imaging (MSEPI) can improve spatial resolution and reduce geometric distortions compared to conventional single-shot EPI by segmenting the acquisition into multiple shots with reduced readout duration and correcting for phase differences between shots during the image reconstruction step [[3], [4], [5]]. However, scan time in MSEPI increases linearly with respect to the number of shots, which limits the number of shots and thus reduction of geometric distortions. Parallel imaging can be used to accelerate acquisition; however, the low baseline SNR of high b-value acquisitions limits the achievable acceleration rates [6]. Compressed sensing is another option to undersample k-space and accelerate the acquisition, however achieving sufficient incoherent sampling with EPI is limited and the low SNR of high b-value images reduced sparsity [7,8].

In recent years, there has been a surge of research on using deep learning (DL) techniques to accelerate MRI data acquisition, particularly to reconstruct images from undersampled k-space data. In the case of DWI, recent studies have employed convolutional neural networks (CNNs) to reduce the number of high b-value acquisitions as an alternative to k-space acceleration [9,10]. These studies have primarily focused on single-shot DWI in the body for all-in-one diffusion or for three diffusion directions. Similar techniques have been used to accelerate multi-shot DWI in the brain [11,12].

This work presents a deep learning denoising technique to accelerate multi-shot DWI in the body, particularly in the rectum, which is more challenging than the brain due the presence of multiple air-tissue interfaces and gas-related motion. Here, number of repetitions acquired is reduced and replaced by the DL algorithm. In this sense the work follows [[9], [10], [11], [12]]. The network is trained using single-shot DWI in the rectum and applied to multi-shot high b-value DWI. The inputs to the network are the accelerated high b-value image (noisy) and the low b-value image (anatomical reference), and the output is the denoised high b-value image. Different acceleration factors are tested on patients with rectal cancer and evaluated quantitatively and qualitatively using scores from expert radiologists. In-depth analysis of the effects of denoising on qualitative scores is performed to assess how deep learning denoising is affecting image features across the scoring criteria.