Back to Feed
Computer Vision / Efficiency & Inference

Evaluating Image Super-Resolution for Brain Lesion Detection

Original: Does FLAIR super-resolution erase or hallucinate small white-matter lesions?
By Zahra Khodakarami, Yue Li, Pulkit Khandelwal, John Detre, Sandhitsu Das, Christopher Brown, David Wolk, Paul Yushkevich Published 2026-08-06

Listen to the summary

Uses a voice available on your device

Playback settings
On this page

Key Takeaways

  • ECLARE outperformed other methods by increasing the Dice score by 0.064 at 3mm and 0.060 at 5mm compared to thick-slice baselines.
  • The primary failure mode across tested super-resolution models is the accidental erasure of small lesions.
  • Erasure rates of lesions increase as the input slice thickness increases.
  • Implicit neural representation models caused a 54% inflation in total lesion volume at 5mm thickness, despite maintaining a low rate of hallucinated features.

Summary & Methodology Analysis

The researchers aimed to determine if super-resolution (SR), a technique using neural models to increase image resolution, maintains diagnostic accuracy for white-matter hyperintensity (WMH) lesions. They processed 29 ADNI subject scans to create a ground truth using 1-mm isotropic HR FLAIR data. They then synthesized thick-slice LR FLAIR by applying Gaussian kernel blurring along the through-plane axis and downsampling to 3 and 5 mm thicknesses. This simulated environment provided a controlled baseline to measure how various upsampling pipelines impact downstream segmentation performance using four existing tools including MARS-WMH, which was selected as the primary reader due to its sensitivity to small lesions.

Interactive System Flowchart

Click diagram to expand and zoom

Cross-Examination & FAQs

A deeper dive clarifying mechanics, constraints, and baseline evaluations.

Q1. What is the primary goal of this research?

To investigate if applying super-resolution to thick-slice MRI scans causes the loss of real small lesions or the creation of false ones.

Q2. Did super-resolution improve segmentation results?

Yes, ECLARE specifically improved the Dice gain by 0.064 at 3mm and 0.060 at 5mm compared to un-reconstructed baseline data.

Q3. What is the biggest risk when using super-resolution for lesion detection?

The primary failure mode identified is the erasure of small lesions, a problem that becomes more frequent as slice thickness increases.

Q4. What techniques were used to upsample the images?

The authors used cubic interpolation, a multi-contrast implicit neural representation (INR) using T1w priors, and a single-contrast self-supervised model known as ECLARE.

Q5. How did the implicit neural representation (INR) model perform?

While it had few hallucinated lesions, the INR model inflated the total WMH volume by 54% when evaluated at a 5mm slice thickness.

Q6. How was the segmentation performance analyzed?

The authors used four segmentation methods (WMH-SynthSeg, segcsvd, MARS-WMH, and TrUE-Net) and focused on detection sensitivity, lesion erasure rates, and hallucination rates.

Q7. Does the study use real clinical thick-slice data?

No, the study utilized simulated thick-slice data generated from 1-mm isotropic HR FLAIR scans rather than scans acquired as real thick-slice FLAIR.

Q8. What are the limitations of the evaluation process?

Limitations include using a single cohort (ADNI), selecting one segmenter post-hoc, and using a single expert's manual mask as reference, which may result in underestimating false positives.

Q9. Which models and tools were evaluated in this paper?

The paper references ADNI, WMH-SynthSeg, segcsvd, MARS-WMH, TrUE-Net, ECLARE, SMORE, ESPRESO, and INR.