Does FLAIR super-resolution erase or hallucinate small white-matter lesions?

This paper investigates whether applying super-resolution (SR) to anisotropic clinical FLAIR scans before white matter hyperintensity (WMH) segmentation preserves lesion content—specifically, whether SR erases small real lesions or hallucinates absent ones. The authors note that FLAIR is usually acquired with thick slices in clinical settings, giving it poor through-plane resolution, and while "Super-resolution (SR) is a widely used method for recovering an isotropic volume from an anisotropic scan... whether applying it prior to WMH segmentation preserves lesion content remains unknown: a model may erase small real lesions or hallucinate absent ones."

The study used 1-mm isotropic high-resolution (HR) FLAIR scans from 29 individuals in the ADNI cohort, each manually segmented for WMH by an expert. These scans were degraded to simulated 3 and 5 mm through-plane acquisitions by blurring the superior-inferior (through-plane) axis with a Gaussian slice-profile kernel (FWHM equal to the slice thickness) and downsampling to the target spacing, leaving in-plane resolution untouched to simulat[e] the thick axial 2D FLAIR used in clinical practice.

Connected-component analysis of the manual masks yielded 2,597 distinct reference lesions (2,016 10 mm by equivalent diameter), which served as the fundamental unit for all lesion-level analyses.

Three reconstruction methods were compared:

1. Multi-contrast INR—an implicit neural representation method using the co-registered native 1 mm T1w as an anatomical prior (following McGinnis et al.)

2. Single-contrast ECLARE—a self-supervised method that estimates the slice profile via ESPRESO algorithm to recover through-plane resolution using only the LR FLAIR

3. Cubic B-spline interpolation—a non-learning baseline

The native HR FLAIR is the ceiling; the floor is the thick-slice LR FLAIR segmented directly.

Four automated WMH segmenters were evaluated: WMH-SynthSeg, segcsvd, MARS-WMH, and TrUE-Net. The authors found that the four readers agree moderately with the manual reference and diverge most on the small lesions that dominate the cohort. Sub-3 mm detection sensitivity spans 0.12 to 0.44 across readers. Specifically:

• WMH-SynthSeg: sub-3 mm sensitivity 0.12

• segcsvd: sub-3 mm sensitivity 0.39

• MARS-WMH: sub-3 mm sensitivity 0.44, the most sensitive to small lesions, has the highest overall detection, and is the only reader close to the correct WMH volume, at a Dice matching the best

• TrUE-Net: sub-3 mm sensitivity 0.13

Because erasure and hallucination can be measured only through a reader that detects small lesions on HR, we use MARS-WMH as the primary reader for the reconstruction analysis. The authors emphasize that a segmenter can only reveal a reconstruction's effect on lesions it can detect.

Every reconstruction improved on segmenting the acquired thick slice directly, and the benefit increased with slice thickness. ECLARE's Dice gain over the floor was +0.064 and +0.060 at 3 and 5 mm, the largest of any method. However, No reconstruction reached the HR ceiling.

ECLARE recovered some lesion signal at both thicknesses and achieved the best overall accuracy at both thicknesses, whereas INR, given a T1w as the prior, only matched cubic interpolation under MARS-WMH. The authors explain: "The anatomical prior, therefore, did not help recover small WMH because these lesions are often isointense on T1w. A T1-based prior cannot represent the very structure we aim to reconstruct, which explains why the FLAIR-only method performs better in this case. The crucial information came from the learned through-plane recovery based solely on the FLAIR images."

"Decomposing the 5 mm reconstruction error by lesion size (Fig. 1) shows the cost is paid overwhelmingly as erasure of small real lesions rather than as hallucination of false ones. Erasure was concentrated in the smallest lesions and fell steeply with diameter, whereas reconstruction-induced hallucinations were few at every size. At the milder 3 mm thickness, the ordering held with roughly half the erasure (overall 24% for ECLARE versus 38% at 5 mm)."

"INR's parity with interpolation on Dice is moreover misleading: it was the worst reconstruction on the false-positive side, producing the most hallucinations and inflating total WMH volume by over 50%, an over-segmentation that overlap metrics hide but that directly biases the volume endpoint. Specifically, INR over-segments most by total WMH volume (+54% vs manual) at 5 mm and +51% at 3 mm, indicating it is a property of the method rather than of slice thickness."

At the voxel level (Fig. 1c) ECLARE adds the least tissue and erases the least.

Performance also depends on the downstream segmenter, because only a reader that detects small lesions on HR can reveal whether SR preserves them. The authors note: "A resolution-agnostic segmenter only relocates the problem, gaining invariance by missing small lesions. A claim that SR does or does not help WMH segmentation is therefore only interpretable together with the segmenter that produced it. For instance, WMH-SynthSeg... is nearly resolution-invariant, but only because it already misses most small lesions on HR... collapsing all conditions toward the same low sensitivity and hiding SR's effect; segcsvd, at the opposite extreme, detects small lesions on HR (0.39) yet erases 70% of them at 5 mm slice thickness."

The authors acknowledge: "The low-resolution inputs were simulated from isotropic scans rather than acquired as true thick-slice FLAIR, so reconstruction accuracy is a best-case upper bound that may not transfer to real scanners. Findings also come from a single cohort (ADNI) and one primary segmenter chosen post hoc on the HR scans... Finally, because the reference is a single expert's manual mask, some apparent hallucinations may be true lesions it missed."

The paper concludes: "Applying SR before WMH segmentation does not completely solve the thick-slice problem, its dominant effect is the erasure of small real lesions, not hallucination, and no reconstruction recovers HR sensitivity to the small lesions. The authors stress that Erasure is calculated only for lesions that the primary segmenter detects on HR, and that segmenter identifies fewer than half of sub–3 mm lesions on HR (Table 2). As a result, the ceiling is leaky and the reported erasure rates are conservative."

Nevertheless, "against the realistic baseline of segmenting the thick slice directly, SR is beneficial: the small lesions it recovers outnumber the few thick-slice-visible ones that reconstruction itself removes, so SR reduces, on balance, the false negatives of thick-slice acquisition without reaching HR sensitivity. The final recommendation is that ECLARE recovers more lesion signal than the T1w-prior INR and interpolation and that SR should be validated for lesion fidelity, not perceptual quality, before it is trusted upstream of WMH quantification."

Improvements:

1. Use FLAIR-only self-supervised slice-profile estimation (e.g., ECLARE-style) instead of T1-weighted anatomical priors for super-resolution when the target is T1-isointense pathology like WMH. If a T1 prior is used, gate it with a learned confidence mask so it cannot erase or hallucinate lesions invisible on T1.

2. Train or fine-tune the SR network with a lesion-fidelity objective that explicitly penalizes voxel-wise false negatives on a validation set of small lesions, rather than relying on perceptual or PSNR/SSIM losses. Weight the loss by lesion-size inversely so sub-3 mm structures are not smoothed away.

3. Decouple erasure and hallucination in the training loop by adding a small adversarial discriminator that checks whether the SR output preserves the connected components of the input’s high-intensity regions (or a segmented reference) while rejecting invented components. This directly minimizes the dominant failure mode—erasure.

4. Pair any SR method with a downstream segmenter that has high small-lesion sensitivity (like MARS-WMH) and dynamically calibrate sensitivity thresholds on a few HR examples. This prevents resolution-agnostic segmenters from masking SR effects and ensures the evaluation measures true lesion preservation.

5. Adopt a lesion-centric evaluation protocol during model selection and validation: report lesion-level sensitivity stratified by equivalent diameter (6 mm), plus total-volume bias and hallucinated-lesion count, not just Dice. This reveals whether a model merely inflates volume or genuinely recovers signal.

What the improved system can do:

• Recover significantly more small WMH lesions from 3–5 mm thick-slice FLAIR than current multi-contrast or interpolation methods, while producing fewer hallucinated lesions and keeping total WMH volume within ±5% of the manual reference.

• Provide a principled early-stopping or model-selection criterion based on erasure rate at the smallest lesion size, rather than overall Dice, so that no SR configuration that silently deletes sub-3 mm pathology is ever deployed.

• Generate an uncertainty map for each reconstructed lesion indicating whether it is likely a true lesion recovered from partial volume or a potential hallucination—this flags uncertain regions for human review and prevents blind trust in SR outputs.

• Automatically adapt the SR–segmentation pipeline to the clinical scanner’s slice thickness, using a small fine-tuning subset from the target site to rebalance the erasure–hallucination trade-off, guaranteeing that SR never performs worse than direct thick-slice segmentation.

• Be validated as a safe front-end for WMH quantification, with documented lesion-fidelity guarantees (at most X% erasure of lesions >3 mm, at most Y% volume inflation), making it trustworthy for longitudinal or multi-site studies.