An arXiv preprint reports that a base nnU-Net using only diffusion-weighted imaging (DWI), without brain extraction, produced a median Dice similarity coefficient (DSC) of 0.84. DSC measures how closely the model's lesion outline overlaps with the reference outline. The same configuration had a median lesion-wise F1-score of 0.77, a median absolute volume difference of 1.37 mL and a median absolute lesion difference of 1. Inference took approximately 5 seconds per case.
A practical test built from three datasets
The underlying question was whether acute ischaemic stroke lesions could be segmented accurately with minimal preprocessing and a clinically feasible inference time. The analysis used retrospectively collected MRI from Nor-COAST, OSCAR and SOOP. Lesions were annotated on high b-value DWI, with derived ADC and FLAIR images available where provided.
Researchers identified 2,309 MRI cases and excluded 129, leaving 2,180 for analysis. Of those, 1,744 were used for training and validation and 436 were held back for testing. The training and validation data were split 80/20. Twelve OSCAR test cases lacking FLAIR were excluded from the comparison with DeepISLES.
The modelling experiment used five-fold cross-validation across four input conditions. The conditions combined brain-extraction status with either DWI alone or DWI plus ADC, and each was assessed with both the base and ResEnc nnU-Net architectures.
Because the performance measures were not normally distributed, the researchers used non-parametric statistical tests. Model comparisons relied on Friedman and Wilcoxon signed-rank tests, while dataset and lesion-volume analyses used Kruskal-Wallis tests, correlation and regression analyses, Bland-Altman assessment, and Bonferroni-corrected testing where appropriate.
The simpler input held its own
The overall differences between input configurations were statistically significant, with p<0.001. But the results did not show a consistent advantage from adding ADC or changing the preprocessing. DSC did not differ between models using DWI alone and those using DWI plus ADC, and no other performance measures differed significantly across the configurations.
Architecture made a more selective difference. ResEnc produced higher DSC than the base model for DWI, DWI plus brain extraction, and DWI plus ADC, with all significant comparisons at p<0.02. The difference was not significant for the combination of DWI, ADC and brain extraction, where p>0.50.
Those gains came with a substantial computational cost. One-fold training took 13 hours for the base model and 55 hours for ResEnc. Approximate inference time was 5.0 seconds per case for the base model and 11.0 seconds for ResEnc.
Against an existing model
Across the three test sets, the base nnU-Net had a median DSC of 0.84, compared with 0.81 for DeepISLES. The difference significantly favored the base model, with p<0.001. Lesion volumes also correlated positively with the ground truth for both systems: the Pearson correlation was 0.978 for the base model and 0.957 for DeepISLES.
The gap was wider in the Nor-COAST subgroup. There, median DSC was 0.81 for nnU-Net and 0.53 for DeepISLES. DeepISLES failed to detect lesions in seven of 51 cases, while nnU-Net failed in one case. This was a dataset-specific subgroup comparison, rather than a test across an entirely unseen population.
Lesion size also mattered in the comparison. DeepISLES tended to produce poorer DSC for lesions measuring 5 mL or less. Its per-case inference took approximately 235 seconds, including brain extraction, versus approximately 5 seconds for nnU-Net.
What the results do not settle
The central limitation is the validation design: the study did not include an out-of-sample cohort unseen by any model during training. The reported results therefore leave performance on a fully external dataset unanswered.
The reference annotations themselves were assessed in OSCAR by two experienced neuroradiologists who independently segmented 70 cases. Their mean inter-rater Dice was 0.86, and disagreements below 0.70 were resolved by consensus.
Within the tested retrospective datasets, the pattern favored a fast DWI-only base model for practical segmentation experiments. ResEnc offered selective DSC improvements at longer training and inference times, while the base model outperformed DeepISLES on the overall DSC comparison. Further testing on cohorts unseen by all models would be needed before broader performance claims could be made.
Paper data and sources
Original title: Deep Learning Segmentation of Diffusion-Weighted MRI Acute Ischaemic Stroke: A Pragmatic Evaluation Across Three Datasets
Authors: Atle Bjørnerud, Till Schellhorn, Thor H. Skattør et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-26
DOI: Not available
Original paper · Full text