MRI data are collected in spatial-frequency space over time, and the reconstruction assumes that the object is the same throughout. When it is not, the inconsistencies between lines of data appear as artefact, whose form depends on the motion and on the order of sampling. Periodic motion during a Cartesian acquisition produces discrete ghosts displaced along the phase-encoding direction; irregular motion produces diffuse blurring and loss of edge sharpness (Zaitsev et al., 2015). Table 1 lists the sources that matter in the abdomen with their usual remedies. One deserves particular note. Transient breathlessness shortly after injection of gadoxetate disodium can spoil the arterial phase of a liver examination, the one phase that cannot be repeated (Davenport et al., 2013).

Existing remedies work in one of three ways. Some stop the motion or wait for it: breath-holding, respiratory triggering and navigator gating. Some make the acquisition less sensitive to it: radial and PROPELLER trajectories sample the centre of k-space repeatedly, so that motion is averaged into mild blurring or streaks in place of coherent ghosts (Pipe, 1999). The rest make the acquisition faster. Parallel imaging with SENSE or GRAPPA uses the spatial sensitivity of receiver coils to replace some phase-encoding steps (Pruessmann et al., 1999; Griswold et al., 2002), and compressed sensing exploits image sparsity to go further (Lustig et al., 2007). Combined with golden-angle radial sampling, these gave rise to free-breathing dynamic imaging and to reconstructions that sort data by respiratory state (Feng et al., 2014; Feng et al., 2016). Each has a cost: noise amplification, residual aliasing, long computation or a longer examination.

Most clinical deep learning reconstructions belong to one family. An iterative reconstruction is unrolled into a fixed number of steps, each alternating a data-consistency operation, which keeps the estimate faithful to the measured samples, with a learned regulariser in the form of a convolutional network. The variational network, the deep cascade and model-based deep learning are the reference designs (Hammernik et al., 2018; Schlemper et al., 2018; Aggarwal et al., 2019). Trained on pairs of fully sampled and undersampled data, such networks suppress noise and aliasing at acceleration factors where parallel imaging fails. Their effect on motion is indirect. They do not model motion; they make it possible to acquire the same image in less time, and a shorter breath-hold is one that more people can manage.

Direct correction is a separate line of work. Image-to-image networks have been trained to map motion-corrupted images to clean ones, and hybrid schemes use a network to guide a physical model of the motion (Küstner et al., 2019; Haskell et al., 2019). These are promising and largely unavailable in clinical products. Both kinds share a known weakness. A learned prior can produce images that look sharp and plausible while departing from the data, and small structures or perturbations unlike anything in the training set can be removed or altered (Antun et al., 2020). The best-known public training set consists of knee and brain examinations (Zbontar et al., 2018), so abdominal performance cannot be assumed from published benchmarks.

The first decision is which motion problem is to be solved, for which patients and on which sequences. The answer should come from the experience the department has of failed examinations. Typical candidates are the breath-held T1-weighted gradient-echo sequence used for dynamic liver imaging, T2-weighted single-shot and fast spin-echo sequences, and diffusion-weighted imaging. A target is stated in operational terms, for example that the dynamic sequence should be acquired in a breath-hold short enough for people who cannot hold for the present duration. A sequence with no identified motion problem is left alone.

Deep learning reconstruction reduces motion artefact only if the acquisition is changed to take advantage of it. The options are a higher parallel-imaging or compressed-sensing factor, fewer signal averages, or partial Fourier sampling, each of which shortens the scan and lowers the signal-to-noise ratio that the network is then asked to restore. Changes are made one at a time from the existing protocol, which is kept unaltered as the reference. Contrast-determining parameters, that is repetition time, echo time, flip angle and fat suppression, are left as they are, so that any change in lesion conspicuity can be attributed to the reconstruction and not to altered weighting. Where a radial free-breathing sequence is available, it is treated as a separate candidate and not mixed with the breath-hold variant (Feng et al., 2014).

Commercial implementations expose few controls, usually a strength or denoising level and sometimes an edge-enhancement option. The setting is chosen on test data before any comparison and then fixed, because adjusting it case by case makes the result irreproducible. The software version, the setting and the sequences to which it is applied are written down; a version update is treated as a new reconstruction requiring at least the technical checks of the next stage. Throughout the evaluation the conventionally reconstructed images from the same raw data are also saved wherever the system allows. They cost nothing in scan time and provide the only direct check on whether a feature seen, or not seen, on the deep learning images is present in the data.

Before clinical use, the candidate protocol is run on a structured phantom to confirm geometric accuracy, the absence of new structured artefact and the behaviour of fine detail and low-contrast inserts at the chosen setting. Global similarity indices such as the structural similarity index are convenient for comparing reconstructions of the same data (Wang et al., 2004), but they summarise the whole image and are insensitive to the loss of a small lesion, so they are recorded without being treated as evidence of diagnostic adequacy. A small number of volunteer examinations, under the arrangements the department has for volunteer scanning, then establishes that the shortened sequence runs as intended and that the breath-hold has in fact become shorter.

The central test is a paired comparison in clinical examinations. Each person undergoes the reference sequence and the candidate sequence in the same session, in an order that is varied to avoid a systematic effect of fatigue or contrast timing. For post-contrast sequences, where both cannot occupy the same phase, the comparison is restricted to phases that can be duplicated, or to consecutive examinations of the same person. At least two radiologists, unaware of the method, score each series on predefined scales for motion artefact, sharpness, noise, artificial texture and overall diagnostic quality, and separately record the focal lesions they detect. Lesion detection against a reference reading matters more than preference scores, because readers tend to prefer smooth images. The sample size, the scales and the analysis are fixed in a written plan, and the work is reported in line with the CLAIM checklist (Mongan et al., 2020).

If the comparison supports adoption, the candidate protocol replaces the reference for the stated indication, and monitoring begins. Two simple measures are proposed. One is the proportion of examinations in which a sequence is repeated or flagged by the reporting radiologist as limited by motion, tracked before and after the change. The other is a log of discrepancies in which a finding on follow-up or on another modality was not visible on the deep learning images, each reviewed against the conventional reconstruction if it was kept. Monitoring is renewed after any software update or change of coil.