An arXiv preprint dated 20 August 2026 reports that AsymFeX, a two-stage model built around brain symmetry, outperformed the compared methods at marking stroke lesions on AISD, a dataset containing 397 NCCT scans. Its Dice overlap score was 0.6796, while HD95, a boundary-distance measure, was 23.53 and absolute volume difference was 7.69. The paper reports statistically significant differences for all three metrics at p < 0.05.
The same design was evaluated on T1-weighted MRI and CT-perfusion data, where it produced competitive scores. But this was a model evaluation using public datasets, not a test of treatment decisions or patient outcomes.
The model compares a scan with its mirror image
The pipeline first aligns each volume to the true mid-sagittal plane, the brain’s central dividing plane. It then processes the aligned volume and its left-right flip through a shared-weight encoder and the AsymFeX module, allowing the system to compare the two sides of the brain.
Its cross-attention component searches a local 3 × 3 × 3 neighborhood, or 27 voxels, for corresponding information rather than using a fully global search.
In a representative profile, local attention reduced attention-related computing work by 531 times, total block work by 28.5 times, peak training memory by 15.6 times and inference time by 11.7 times compared with global attention, while parameter counts stayed the same. The profile used one resolution and batch size, and no variability estimate was reported.
A mixed set of tests
The evaluation used three public datasets. AISD supplied 397 NCCT scans, split into 345 training scans and 52 test scans. ATLAS v2.1 included 1,271 T1-weighted MRI volumes from 44 research cohorts; the study used 655 public training volumes and evaluated 300 test volumes with released ground-truth masks.
The ISLES’24 result needs particular care. Its official 96-case test split was withheld, so the study instead used 149 available cases, divided into 109 training cases and 40 test cases. It therefore does not report a result on that official withheld split.
The clearest gain came on AISD
On AISD, the model’s Dice score was 0.6796. Its HD95 and absolute volume difference were 23.53 and 7.69, respectively, and the comparison with baselines was statistically significant across all three measures at p < 0.05. Dice describes overlap between predicted and reference lesion regions, while HD95 and absolute volume difference describe boundary and volume error.
An ablation test, in which parts of the design were removed, was associated with worse reported performance. Across three configurations, Dice changes were 8.1%, 5.6% and 4.9%, while HD95 changes were 34.5%, 52.27% and 27.79%; all differences were reported at p < 0.05.
Using the study’s 70 mL cutoff for the small-lesion analysis, Dice was 0.6131. Predicted and reference lesion volumes had a Pearson correlation of 0.9854 with p < 0.001. The mean volume difference was −5.68 mL, and the reported Bland–Altman limits of agreement ran from −33.30 to 21.93 mL.
Performance depended on the scan inputs
On ATLAS T1-weighted MRI, Dice was 0.6353 ± 0.2597 and HD95 was 20.9578 ± 23.4676. On ISLES’24, Dice was 0.6354 ± 0.2520 for infarct and 0.6339 ± 0.2348 for penumbra, with HD95 values of 15.1846 ± 17.2990 and 13.8642 ± 13.7755, respectively.
The ISLES’24 NCCT-only test was near-failing across methods: Dice was 0.1766 for nnU-Net, 0.1763 for Attention U-Net and 0.1786 for the proposed model, with HD95 between 159 and 163 mm. When perfusion maps were included, the proposed model’s Dice was 0.6354. This large difference applies to the reported input settings and does not establish that NCCT performs poorly at every stroke stage.
The study also tested whether uncertainty could flag errors
To test whether the model could identify suspect regions, the authors used Monte Carlo dropout: 30 stochastic passes of the model. AUROC, a measure of how well uncertainty separates errors from non-errors, ranged from 0.7053 to 0.8102 across AISD, ATLAS and the two ISLES’24 targets. AUPRC exceeded prevalence for every dataset.
Calibration, which asks whether a model’s confidence matches its actual accuracy, was strongest on AISD. Expected calibration error was 0.0807 on AISD, compared with 0.1437 on ATLAS, 0.1694 for ISLES’24 infarct and 0.1660 for penumbra. The study did not validate a clinical decision threshold, so it has not shown that these uncertainty scores can tell clinicians when to act.
Promising model evaluation, not clinical evidence
The symmetry assumption sets an important boundary on the findings. The analysis notes that substantial midline shift, encephalomalacia or ventricular enlargement unrelated to the acute event could disrupt the left-right comparison on which the method relies. Independent prospective external validation was not reported.
The manuscript is labeled arXiv:2608.19769v1 and dated 20 August 2026. Its results support comparative performance on the reported public datasets, but they do not show improved treatment selection, patient outcomes or prospective clinical effectiveness.
Paper data and sources
Original title: AsymFeX: A Symmetry-Driven Framework for Ischemic Stroke Segmentation Across Imaging Modalities and Stroke Stages
Authors: Maunil Shah, Vaanathi Sundaresan
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-20
DOI: Not available
Original paper · Full text