Generative Adversarial Networks for MRI Artifact Reduction

That distinction is the central problem in validating generative adversarial networks for MRI artifact reduction. MRI reconstruction is not a cosmetic operation. The network modifies an inverse problem governed by k-space sampling, coil sensitivity, gradient performance, motion state, noise statistics, and sequence contrast. It can suppress visible ghosting while also attenuating a lesion boundary, shifting an edge, or fabricating texture that was never encoded by the acquisition.

The evaluation therefore has to operate on several layers. Full-reference signal metrics measure numerical similarity. Perceptual metrics estimate visual plausibility. Focus measures test edge retention. Clinical grading determines whether the output remains interpretable. Downstream task performance asks the harder question: does the corrected image preserve the information required for segmentation, registration, lesion detection, or quantitative measurement?

No single score answers all five.

The first constraint: MRI artifact reduction is an inverse problem

The corrupted image is not merely a clean image with noise painted over it. Motion and undersampling alter the measured signal before reconstruction. In simplified form, the scanner acquires samples according to an encoding operator that includes the spatial trajectory through k-space, coil-dependent sensitivity, gradient timing, and sequence-specific contrast. A reconstruction network then estimates an image from incomplete or inconsistent measurements.

That estimate is conditional. It is constrained by the training distribution.

A GAN can learn that certain anatomical structures usually appear in particular spatial relationships. This prior helps when the artifact is represented in the training data. It becomes dangerous when the acquisition violates those assumptions. A motion pattern from a head scan is not interchangeable with respiratory motion in cardiac MRI. Phase-encode ghosting is not equivalent to Gibbs ringing. Parallel-imaging noise amplification does not behave like random Gaussian noise. The artifact mechanism determines what information is absent, displaced, or corrupted.

The validation protocol must preserve that distinction.

A test set assembled from synthetic image degradation answers one question: can the model invert the artificial corruption applied by the investigator? A set of prospectively motion-corrupted clinical scans answers another: does the network tolerate the physics and anatomy of real acquisition? The two results should not be merged into one headline number.

A higher similarity score proves that the output resembles the reference. It does not prove that the network preserved the lesion, boundary, or quantitative contrast that the reference was meant to represent.

Before calculating SSIM or PSNR, define the corruption regime. Record the anatomical region, pulse sequence, field strength, acceleration factor where applicable, motion type, k-space manipulation, coil configuration, and reconstruction pathway. If these variables are hidden, the metric is not reproducible. It is an ornament attached to an undocumented pipeline.

SSIM and PSNR: useful baselines with narrow jurisdiction

In medical image-to-image translation studies involving GANs and CNNs, SSIM appears in 84% of evaluated literature, while PSNR appears in 61%. Their popularity is justified. Both are easy to compute, easy to compare, and widely recognized. Neither should be treated as a clinical endpoint.

PSNR measures error energy

Peak Signal-to-Noise Ratio is derived from mean squared error between the reconstructed image and a reference image. In simplified terms, lower pixelwise error yields higher PSNR. The unit is decibels.

PSNR is sensitive to intensity deviation. It penalizes a voxel-by-voxel mismatch even when the mismatch is visually negligible. It can also reward excessive smoothing. If a network removes high-frequency structures and produces a conservative, blurred estimate, the mean squared error may fall. The resulting PSNR can improve while spatial resolution degrades.

The reference must therefore be genuinely paired with the corrupted input. A clean image from a different acquisition is not a valid pixelwise ground truth. Changes in patient position, contrast timing, slice prescription, coil loading, or physiological state contaminate the comparison. The network may be penalized for harmless registration differences or rewarded for reproducing broad anatomy while missing fine pathology.

Report the intensity normalization procedure. A PSNR value without a defined dynamic range is incomplete. The same reconstruction can yield different PSNR values after min–max scaling, z-score normalization, percentile clipping, or vendor-specific intensity correction.

SSIM measures local structural agreement

Structural Similarity Index Measure compares local luminance, contrast, and structural patterns. It is less crude than a pure pixelwise error metric. It can tolerate certain small intensity shifts while responding to local structural changes.

That tolerance is also a limitation. SSIM is computed over windows. A small but clinically decisive abnormality can be diluted by surrounding normal tissue. A model can preserve large-scale anatomy and achieve a high SSIM while altering a thin cortical lesion, vessel wall, hippocampal boundary, or postoperative margin.

SSIM is also affected by window size, dynamic range, and implementation details. State those parameters. Report whether the metric is calculated on the full field of view, a brain mask, tissue-specific regions, or artifact-affected regions only. A global score can conceal local failure.

Head MRI motion-artifact studies using conditional GANs have reported SSIM above 0.9 and PSNR above 29 dB. These are strong results within the reported experimental conditions. They are not an industry-wide acceptance threshold. The correct interpretation is conditional: the model met those values on that modality, anatomy, corruption distribution, and reference construction.

Read the metrics together

A useful baseline report should include at least:

• PSNR with the intensity range and normalization method.
• SSIM with windowing and region-of-interest definition.
• Mean absolute error or another robust voxelwise error measure.
• Metrics calculated separately for artifact-heavy and artifact-light regions.
• Confidence intervals or dispersion across subjects, not only a cohort mean.
• Performance stratified by artifact severity and acquisition condition.

The last point is routinely neglected. A mean PSNR can improve because the model performs well on mild corruption. That does not establish reliability on the severe cases for which artifact reduction is clinically relevant.

The table is not a ranking. It is a map of jurisdiction. Each metric governs a different failure surface.

Perceptual quality: LPIPS and FID are secondary evidence

PSNR and SSIM require a reference image. Clinical MRI often lacks a perfectly paired, artifact-free acquisition. This is where perceptual metrics enter. LPIPS, or Learned Perceptual Image Patch Similarity, compares deep feature representations rather than raw intensity values. FID, or Fréchet Inception Distance, compares distributions of feature embeddings across generated and reference image sets.

They answer questions that pixel metrics cannot answer. Does the output occupy a plausible image manifold? Does it preserve broad visual characteristics across a cohort? Does the reconstruction look structurally coherent when no direct clean reference exists?

They do not answer whether the output is physically faithful.

LPIPS is generally inherited from computer-vision feature spaces. A representation trained on natural images does not automatically encode MRI-relevant distinctions. T1-weighted contrast, diffusion signal, susceptibility effects, fat suppression, and subtle cortical abnormalities have no obligation to align with the feature geometry learned from photographic data. A low LPIPS distance can coexist with a clinically meaningful intensity error.

FID has a different problem. It is a population-level statistic. It can identify a distributional shift between generated and reference images, but it does not expose the single case in which a small lesion disappears. Its value depends on the feature extractor, sample size, preprocessing, and anatomical homogeneity of the evaluated cohort. A score computed across mixed sequences or anatomical regions becomes difficult to interpret because the distribution contains multiple sources of variation.

If LPIPS or FID is reported, disclose:

1. The feature extractor and its training domain.

2. Whether slices, volumes, patches, or projections are used.

3. The preprocessing and intensity standardization pipeline.

4. The number of subjects and independent acquisitions.

5. Whether the metric is calculated globally or by sequence and anatomy.

6. How outliers are inspected rather than hidden inside a cohort average.

A GAN can optimize perceptual plausibility. That is precisely why perceptual metrics require adversarial scrutiny. The discriminator rewards outputs that look like valid MRI. It does not possess a reliable concept of patient-specific truth unless the training objective and data constraints enforce one.

This is the difference between plausible reconstruction and faithful reconstruction. The former is an image property. The latter is a measurement property.

Clinical grading must be structured, not ceremonial

Visual review by radiologists is not an optional decorative layer added after the mathematics. It is the mechanism that tests whether the mathematical improvement survives contact with anatomy.

A five-point Likert scale is commonly used for subjective assessment. The scale can grade artifact severity, overall image quality, anatomical sharpness, diagnostic confidence, or preservation of pathology. Those are not interchangeable endpoints. The rubric must define them separately.

A reviewer who assigns a higher score because ghosting is reduced may still be unable to determine whether a small lesion has been altered. The grading form should therefore force attention to the structures most vulnerable to generative correction:

• Lesion conspicuity and margin definition.
• Cortical and ventricular boundaries.
• Vessel continuity and caliber.
• Tissue contrast relevant to the sequence.
• Residual motion ghosting.
• New texture, ringing, or edge structures.
• Diagnostic confidence for the intended clinical task.

In simulated cardiac MRI motion-artifact reduction, reported PSNR improved from 23.85 ± 2.85 dB to 27.91 ± 1.74 dB, while SSIM increased from 0.71 ± 0.08 to 0.83 ± 0.05. Subjective grading on a five-point Likert scale also improved, from 2.44 ± 1.08 to 4.44 ± 0.66 for simulated artifacts. For real-world clinical artifacts, the reported visual score increased from 3.03 ± 0.73 to 3.73 ± 0.87.

These results are coherent. The numerical and visual measures move in the same direction. They still do not establish diagnostic safety. The simulated and real-world cohorts represent different validation environments. The magnitude of improvement is not transferable without matching the artifact mechanism, sequence, anatomy, and review protocol.

The grading study itself needs controls. Use blinded, randomized presentation where feasible. Keep the display conditions consistent. Separate training cases from evaluation cases. Report reader count, expertise, inter-reader agreement, and whether readers saw the corrupted and corrected images side by side. Side-by-side review improves detection of differences but can also inflate perceived benefit. Independent review is closer to deployment. Both can be informative if labeled.

The clinical question should be explicit. If the intended use is triage, grade whether the corrected image preserves the findings required for triage. If the purpose is automated segmentation, visual quality alone is insufficient. If the output will support quantitative MRI, intensity fidelity becomes a primary endpoint rather than a secondary aesthetic concern.

Tenengrad: sharpness is not truth

The Tenengrad Focus Measure estimates image focus through gradient energy, commonly using a Sobel operator. Strong local gradients produce a higher score. In MRI artifact reduction, this makes Tenengrad useful for tracking edge preservation and detecting over-smoothing.

Its interpretation is narrow.

A blurred reconstruction loses gradient energy. A sharpened reconstruction gains it. But ringing, noise amplification, and hallucinated boundaries also increase gradient energy. Tenengrad cannot determine whether the edge belongs to the patient or to the reconstruction model.

This matters because GANs can generate high-frequency content that appears anatomically convincing. A higher focus measure may indicate genuine restoration of a boundary. It may also indicate an aggressive prior imposed by the generator. The metric should therefore be calculated in anatomically defined regions and paired with visual review. A global increase is weak evidence. A stable or improved value around known boundaries, without new false edges in homogeneous tissue, is more informative.

The same principle applies to other sharpness measures. They quantify image behavior, not anatomical provenance.

A practical evaluation separates at least three regions:

• Artifact-dominant zones, where ghosting or blur is expected.
• Boundary zones, where edge preservation matters.
• Relatively homogeneous tissue, where artificial texture and noise amplification should be penalized.

The model should yield lower artifact energy without uncontrolled gradient inflation. If Tenengrad rises sharply while SSIM remains stable and radiologists report new edge structures, the network is not restoring information. It is changing the image prior.

Design the benchmark around failure modes

A credible GAN evaluation begins with the split, not the score. Patient-level separation is mandatory. Slice-level randomization can place adjacent slices from one subject in both training and test sets. That creates leakage through anatomy, intensity profile, and acquisition characteristics. The result is a validation number with degraded scientific value.

The benchmark should include multiple corruption sources:

1. Synthetic motion or artifact simulation.

This provides paired references and enables PSNR, SSIM, and error-map analysis. It is controlled and useful. It is also limited by the realism of the simulation.

2. Retrospective clinical artifacts.

These reflect actual acquisition failures, but a clean paired reference is usually absent. Evaluation must rely more heavily on expert grading, task performance, and consistency analysis.

3. Prospective or independently acquired data.

This tests whether the model tolerates protocol variation, scanner differences, patient motion patterns, and changes in reconstruction software.

4. Severity-stratified cases.

Mild, moderate, and severe corruption should not be collapsed. A model that improves mild motion and fails severe motion requires a different deployment claim.

5. Out-of-distribution acquisitions.

Change the sequence, anatomy, field strength, acceleration pattern, or vendor reconstruction where possible. The purpose is not to punish the model. It is to measure the boundary of its learned prior.

The generator’s input and output must also be audited in the measurement domain. If raw k-space is available, compare data consistency before and after correction. If the model operates only on reconstructed images, state that limitation. Image-domain correction cannot be assumed to preserve acquisition-domain constraints.

For quantitative MRI, add parameter-map stability. Artifact reduction that changes T1, T2, ADC, susceptibility, or perfusion estimates can be unacceptable even when the image looks cleaner. The output must be evaluated for bias, repeatability, and spatial consistency. A visually improved map with altered absolute values is not a successful correction.

Downstream tasks reveal what image metrics conceal

The strongest test is often not whether the corrected image looks better, but whether an independent task performs better without acquiring false confidence.

For segmentation, compare Dice score, Hausdorff distance, boundary error, and lesion-volume bias. For registration, measure alignment error and the stability of estimated transformations. For lesion detection, report sensitivity and false-positive behavior across lesion size and artifact severity. For radiomics, assess whether extracted features remain stable under correction and whether their associations with the target clinical endpoint change.

The evaluator should not use the same model family for both correction and verification without controls. A segmentation network trained on GAN-corrected images may simply learn the generator’s artifacts. Test with an independent model, conventional reconstruction, and expert annotations where available.

A useful design includes four image conditions:

• The original artifact-corrupted image.
• The GAN-corrected image.
• A reference or clinically accepted reconstruction.
• A control reconstruction using a non-generative method.

The control matters. If a conventional denoiser or physics-constrained reconstruction yields the same downstream improvement with less alteration of anatomy, the GAN has not demonstrated a unique advantage. It has demonstrated one route to the endpoint.

Artifact suppression is not the endpoint. Preservation of the clinical task is the endpoint.

Failure analysis should be case-based, not limited to aggregate means. Inspect the cases with the largest disagreement between metrics. High SSIM with poor radiologist grading is a critical category. So is low PSNR with stable clinical interpretation. These disagreements identify whether the evaluation is measuring the wrong property or whether the model is exploiting a weakness in the reference construction.

A compact validation architecture

A defensible report can be organized as a layered matrix rather than a single leaderboard:

The word minimum is not the point. The separation of claims is.

A model can pass signal fidelity and fail clinical review. It can pass perceptual quality and fail quantitative mapping. It can improve segmentation while degrading radiologist confidence. Each pattern identifies a different engineering problem. Collapsing them into one composite score removes the information required to fix the system.

What a credible conclusion looks like

The final claim should match the experiment. If the study uses simulated motion in cine CMR with paired references, claim improvement under simulated motion. If it includes real-world artifacts and blinded expert review, claim improved visual quality under those acquisition conditions. If it evaluates lesion detection or segmentation independently, claim downstream task behavior. Do not expand a narrow result into a general statement about MRI.

The unknown remains decisive: there is no universal PSNR or SSIM cutoff applicable across all MRI modalities and anatomical regions for GAN validation. Thresholds are sequence-dependent, anatomy-dependent, and task-dependent. A PSNR above 29 dB may be meaningful in one head MRI experiment and irrelevant in another. An SSIM above 0.9 may coexist with clinically important local alteration.

The correct question is not merely how to check generative adversarial networks for MRI artifact reduction. It is how to determine which information the network removes, preserves, or invents under a defined acquisition regime. That requires metrics, but metrics are only instruments. Their jurisdiction ends where patient-specific anatomy begins.

A GAN is acceptable when its correction is numerically defensible, perceptually controlled, physically constrained, clinically reviewed, and useful for the intended downstream task. Anything less is image enhancement with an incomplete audit trail.