Source-linked AI summary
Adaptive Post-Processing Drives Instance-Level Detection in Stroke Lesion Segmentation
Qinghui Liu, Jon André Ottesen, Atle Bjørnerud, Kyrre Eeg Emblem
TL;DR
Most stroke lesion-segmentation pipelines optimize voxel overlap, leaving instance-level detection—especially for small lesions—underaddressed. This paper introduces volume-conditioned adaptive post-processing and finds it improves Lesion-F1 by 0.032, roughly six times more than tested architectural changes.
Problem
Voxel-overlap training and evaluation can underrepresent instance-level detection, particularly when small-lesion predictions fall short of the IoU ≥0.25 matching gate.
Method
VCAP adapts component-size and binarization thresholds to predicted lesion burden, with thresholds jointly optimized on pooled out-of-fold predictions.
Results
A 0.032 unbiased cross-fold Lesion-F1 gain was roughly six times larger than any tested architectural change, while Viola2Plus raised small-lesion detection by 3.7 points with flat small-lesion Dice.
Takeaways & Limitations
Instance-aware metrics can reveal detection improvements that Dice alone misses, supporting direct examination of detection rate and volume bias.
Takeaways & Limitations
The reported 5-fold internal estimate is not an official test-set result.
Abstract
from arXiv · showhide
Instance-level lesion detection has been an increasingly larger focal point in medical image segmentation besides the more standard voxel-level overlap. Still, most pipelines are trained and post-processed for voxel overlap alone. In particular, the mismatch is most pronounced for small lesions, where a near-miss prediction---substantial overlap that falls just short of the instance-matching threshold---scores the same as a complete miss. In our ISLES'26 submission, we found that closing this gap mattered far more in post-processing than in architecture design. Our Volume-Conditioned Adaptive Post-Processing (VCAP) scheme adjusts component-size thresholds to each case's predicted lesion burden, improving Lesion-F1 by 0.032 (unbiased cross-fold estimate)---approximately 6 times larger than any architectural change we tested. A resolution-aware attention architecture (Viola2Plus), designed for small-lesion segmentation, shows why the distinction matters: it left small-lesion Dice unchanged but raised small-lesion detection rate by 3.7\%, a real effect voxel-overlap metrics alone would have missed. Under 5-fold cross-validation on the 1,453-case training set, our post-processed two-architecture ensemble achieves Dice 0.651 and Lesion-F1 0.614, versus 0.644 and 0.573 for the unprocessed single-model baseline.
1 Introduction
Voxel-overlap training can fail at the instance-matching gate, especially for small lesions and near-miss predictions. The paper therefore focuses on predicted lesion-burden-calibrated post-processing, while examining whether attention architecture can improve small-lesion detection without changing voxel overlap.
- Motivation: Voxel-overlap losses are dominated by large lesions, leaving small-lesion performance underrepresented during training.The passage contrasts a 300 ml lesion with a 0.05 ml lacunar lesion in gradient contribution.
- Evaluation: The evaluation includes Dice, instance-matched Lesion-F1, Absolute Lesion-count Difference, Absolute Volume Difference, and PR-AUC.Lesion-F1 uses an IoU ≥0.25 instance-matching criterion.
- Motivation: If no predicted component reaches IoU ≥0.25 with a ground-truth lesion, instance-matched Dice and Lesion-F1 are forced to zero despite voxel overlap.This instance-matching gate was confirmed at the evaluation-library level.
- Approach: The paper targets the gap between predicted lesion burden and scored lesion burden using post-processing calibrated to predicted lesion burden.The resulting configuration transfers across architectures without retraining.
- Contribution: Attention architecture can raise small-lesion detection without changing voxel overlap, according to stratified failure analysis.This result motivates distinguishing instance-level detection from voxel-level overlap.
2 Methods
The methods use heterogeneous multi-center MRI data, controlled comparisons between a plain nnU-Net and Viola2Plus, and Volume-Conditioned Adaptive Post-Processing (VCAP) to adapt component filtering to predicted lesion burden. VCAP combines burden-dependent component thresholds with recall-first binarization and an empty-prediction rule.
- Data and preprocessing: 1,453 multi-center T1-weighted MRI cases from 72 acquisition cohorts form the training set, with substantial variation in orientation, spacing, lesion volume, metadata completeness, and intensity scale.Time since stroke onset is missing for 23.5% of cases, and lesion volume has a 5.2 ml median with 5 empty ground-truth cases.
- Architecture comparison: Architecture comparisons vary only the network, using identical geometry, data, fold partitions, 1 mm isotropic spacing, patch size, batch size, and deep supervision.The baseline is a stock nnU-Net PlainConvUNet, while Viola2Plus adds tri-axial global attention, high-resolution local spatial gates, and deep decoder stages.
- Architecture comparison: Attention multipliers start as the identity, while the local gate is boost-only and bounded to; both models use Dice plus cross-entropy loss.This initializes Viola2Plus from the plain baseline and permits amplification without suppression.
- Adaptive post-processing: VCAP makes the minimum connected-component volume a function of each case’s predicted lesion burden, addressing the incompatible behavior of any single global filter.A 0.02 ml filter preserves lacunar infarcts but retains false positives near large lesions, whereas 0.1 ml removes noise but discards small true positives.
- Adaptive post-processing: Below 35 ml predicted burden, VCAP uses a permissive 0.02 ml filter; above 35 ml, it switches to a stricter 0.1 ml filter for noisy fragments.A near-empty rule zeroes retained predictions below 0.02 ml, and binarization uses a threshold of 0.35 rather than 0.5.
3 Results
Across pooled 5-fold cross-validation, VCAP produced the paper’s largest practical gains in lesion-level performance, while Viola2Plus improved small-lesion detection without improving Dice. The results therefore distinguish instance retrieval from voxel-level overlap.
- VCAP results: 0.261 → 0.667 was the case-level Lesion-F1 change after VCAP, while spurious components fell from 17 →5 and ALD from 14 →2.The example illustrates VCAP’s removal of spurious components in a thick-slice case.
- VCAP results: 0.032 was VCAP’s unbiased Lesion-F1 gain over the unfiltered baseline, stable across folds under leave-one-fold-out nested selection.The full-data estimate was 0.037, but 0.032 was treated as the honest effect size.
- Small-lesion performance: 3.7 points was Viola2Plus’s detection-rate increase across 188 small-lesion cases, from 0.713 to 0.750.The improvement was about three times the architecture’s cohort-wide detection improvement.
- Small-lesion performance: +6.3% points was Viola2Plus’s detection gain in the T1a tier, while small-lesion Dice remained flat or slightly degraded.Detection increased across all volume tiers, and PR-AUC improved for T1b and T2 lesions.
4 Discussion and Conclusion
VCAP produced a larger instance-level gain than any tested architecture change, while small-lesion detection improved without changing overlap. These conclusions are tempered by limited empty-ground-truth data, increased AVD, and the absence of an official test-set evaluation.
- Discussion: 0.032 Lesion-F1 gain from VCAP was roughly 6 times larger than the 0.006 gain from any architecture change tested.Ensembling, post-processing, and architecture gains were largely independent and stacked rather than competed.
- Discussion: Small-lesion architecture changes raised detection while leaving overlap flat, an effect Dice-only evaluation would miss.Instance-aware claims should be checked against detection rate and volume bias directly.
- Limitations: 5 of 1,453 cases belonged to the empty-ground-truth subgroup, limiting strong conclusions about that subgroup.The paper also notes that the 5-fold internal estimate is not an official test-set result.
- Limitations: 0.07 ml AVD increase was the cost of lowering the binarization threshold to rescue borderline lesions.This trade-off accompanied VCAP’s gain.