médical IA

AI super-resolution in brain MRI: does reconstructing thin slices erase, or invent, small white-matter lesions?

Published on August 8, 2026 · 12 min read

Eight researchers at the University of Pennsylvania ask an unusual question: when you "enhance" a thick-slice brain MRI with artificial intelligence before automatically counting its white-matter lesions, does that enhancement preserve the lesions, remove them, or invent new ones? On 29 high-resolution FLAIR scans from the ADNI cohort, expert-segmented by hand and then artificially degraded, they show that the dominant danger of super-resolution is not hallucination — inventing an absent lesion — but the silent erasure of real small lesions, a phenomenon that worsens as slices get thicker. A key nuance: every reconstruction still beats the raw thick slice, a self-supervised model (ECLARE) preserves signal best, and a supposedly sophisticated method does no better than classical interpolation. The study is a valuable methodological warning, limited by a tiny sample and only-simulated degradation.

The context

White matter hyperintensities (WMH) are bright patches that appear on a type of brain MRI called FLAIR. The FLAIR sequence (Fluid-Attenuated Inversion Recovery) is a way of acquiring the image that suppresses the signal from cerebrospinal fluid to make tissue lesions stand out. These hyperintensities are a marker of cerebral small-vessel disease and are associated with the risk of stroke, cognitive decline, and neurodegenerative disease. Quantifying them — measuring their number and volume — is used in research and in neurology clinical trials.

The problem: in clinical practice, FLAIR is often acquired in thick slices, 3 to 5 mm. The image is well resolved within the slice plane but coarse along the perpendicular axis (so-called through-plane resolution). This is an anisotropic acquisition: the voxels — the three-dimensional "pixels" — are not cubic. Yet small lesions, the ones that matter most for spotting early involvement, can be diluted between two thick slices and become undetectable.

The common fix is super-resolution (SR): an algorithm reconstructs an isotropic volume (cubic 1 mm voxels) from the coarse acquisition, to make the image finer before analysis. But no one had checked the question that really matters: does this reconstruction preserve the lesion content? A model can erase a real small lesion, or fabricate one that does not exist. That is precisely the gap this paper fills.

The method

The design is that of a controlled experiment. The authors start from 29 high-resolution FLAIR scans (1 mm isotropic) from the ADNI cohort (Alzheimer's Disease Neuroimaging Initiative, a large public neuroimaging database), each manually segmented for WMH by an expert — this is the ground truth. They then artificially degrade each volume to simulate thick-slice acquisitions of 3 mm and then 5 mm. For a single brain, one thus has both the ideal version (high resolution) and the degraded version, enabling a lesion-by-lesion comparison impossible in real life, where you never have both.

Three methods then reconstruct the fine volume from the degraded one. The first is a multi-contrast implicit neural representation (INR): a neural network learns to describe the image as a continuous function of spatial coordinates rather than as a grid of voxels. The second, ECLARE, is a single-contrast self-supervised model — "self-supervised" meaning it learns from the image itself, without labels provided by a human. The third is cubic interpolation, a classical mathematical method, without learning, serving as an honest baseline: if the AI does no better, its value is doubtful.

To judge, the authors bound the problem. Segmentation obtained directly on the thick slice (without reconstruction) sets the floor; segmentation on the original high-resolution image sets the ceiling. Each reconstruction is then assessed between these two bounds. Because the measurement depends on the tool counting the lesions, they consider four automatic WMH segmenters (WMH-SynthSeg, segcsvd, MARS-WMH, TrUE-Net) and run the main analysis with the one most sensitive to small lesions on high-resolution images, MARS-WMH — a conservative choice that gives super-resolution its best chance. Three dedicated metrics: detection sensitivity (share of true lesions recovered), erasure rate (lesions present at high resolution but lost after reconstruction), and hallucination rate (predicted objects absent from both the manual and the high-resolution segmentation).

The results

The central result is clear and counter-intuitive. The dominant effect of super-resolution is not hallucination — the usual fantasy about generative models "inventing" lesions — but the erasure of real small lesions. In other words, by making the image prettier, the reconstruction wipes out real signal more often than it fabricates any. And this erasure increases with slice thickness: the coarser the starting acquisition (5 mm rather than 3 mm), the more small lesions disappear.

A second, more reassuring lesson: every reconstruction improves detection compared with the raw thick slice. Doing something is better than doing nothing; the floor really is the worst of worlds. A third lesson, decisive for anyone choosing a tool: among the three methods, the self-supervised model ECLARE best recovers the signal of small lesions, at both thicknesses tested, while the implicit neural representation does no better than cubic interpolation. A method's advertised sophistication therefore does not guarantee its clinical fidelity.

Translated into concrete stakes, without over-reading numbers the preprint reserves for its tables: if a step of AI super-resolution is inserted before measuring a patient's lesion burden, one risks underestimating that burden, and this bias strikes precisely the small lesions — those of early involvement, the very ones one would rather not miss. In a screened population, this could shift patients toward a lower severity category than reality. But this figure describes an effect measured on simulated data, not proof of clinical impact.

What is good

The right question, and a protocol built to answer it. Rather than celebrating a flattering image metric (sharpness, visual likeness), the authors go after what matters downstream: are the lesions preserved? The floor/ceiling design, the high-resolution ground truth with expert manual segmentation, and the lesion-by-lesion analysis make the measurement interpretable. Coining two explicit metrics — erasure rate and hallucination rate — provides a reusable vocabulary for auditing any reconstruction module.

An honest, nuanced result, against the slogans. The message is neither "AI hallucinates, beware" nor "super-resolution fixes everything." It is a precise middle ground: the real risk is silent erasure, it worsens with thickness, but reconstructing remains preferable to the raw slice. This absence of posturing — neither doom nor triumph — is exactly what one expects from a useful evaluation.

A non-AI baseline that reframes the claims. By including cubic interpolation as a comparator, the study avoids the classic trap where any learning method looks good for lack of a simple baseline. The verdict — INR does not beat cubic, only ECLARE adds value — would not have been possible without this safeguard. The ADNI data are public and the evaluated methods are existing, identifiable tools, which makes the work verifiable.

What is less good

A tiny sample, without external validation. Twenty-nine subjects, all from ADNI — a cohort oriented toward Alzheimer's disease, hence elderly and unrepresentative of the full range of situations in which WMH are measured. With so few brains, the estimates are fragile and nothing guarantees transfer to other scanners, other populations, other protocols. This is the classic population bias, here compounded by size. The authors, for that matter, present their work as a pilot workshop study, not a validation.

Simulated degradation, not real thick slices. The 3 and 5 mm acquisitions are not real clinical thick-slice MRIs: they are down-sampled versions of a fine volume. Yet a genuinely thick acquisition carries other imperfections — noise, motion artefacts, contrast differences — that the simulation does not faithfully reproduce. The gap measured here is therefore a proxy: in real life, the erasure could be more pronounced, or different. This is a limit only validation on real data would lift.

A conclusion dependent on the measurement chain, with no human reader. The erasure and hallucination rates depend on the pair {super-resolution method × automatic segmenter}. The main analysis rests on a single segmenter, MARS-WMH, chosen for its sensitivity; other tools would give other figures. Above all, the whole pipeline is automatic: no radiologist re-reads the reconstructed images to say whether the erased lesions would have been seen by a human eye. What is measured is thus a machine-versus-machine effect, useful but not predictive of what a clinician would see.

What it changes

For the research community, the contribution is methodological and immediate: a super-resolution module should not be evaluated on generic image metrics (sharpness, PSNR, SSIM) but on the preservation of clinically relevant downstream content. The erasure and hallucination rates proposed here offer a framework transferable to other tasks — tumour, plaque, or micro-bleed segmentation — where a "beautifying" reconstruction could destroy signal. The obvious next step is replication on real data, at larger scale, with several segmenters and human oversight.

For clinicians and radiologists, nothing to deploy today, but a caution to integrate: inserting an AI super-resolution step before WMH quantification — in a lesion-burden study, a neurological trial, a longitudinal follow-up — is not neutral and can bias the measurement downward, especially for small lesions and the thickest acquisitions. The practical corollary remains that, for want of better, reconstructing beats relying on the raw thick slice, and that between methods, the choice matters.

For patients and the public, the lesson is legible and holds beyond the brain: "enhancing" a medical image with AI is never free. A sharper, more pleasant-to-look-at reconstruction can, in the same motion, wipe out a real detail. Visual sharpness and diagnostic fidelity are not the same thing. This study does not say to reject these tools — they help — but never to mistake a prettier image for a truer one.