Adaptive Post-Processing Drives Instance-Level Detection in Stroke Lesion Segmentation

Details

Researchers from the ISLES'26 competition find that the bottleneck for instance-level stroke lesion detection is post-processing strategy, not model architecture. Their Volume-Conditioned Adaptive Post-Processing (VCAP) scheme adjusts component-size thresholds dynamically based on each scan's predicted lesion burden, yielding a Lesion-F1 improvement of 0.032 — approximately 6 times larger than any architecture change tested. The key insight is the mismatch between training objectives and evaluation metrics: most pipelines optimise for voxel-level overlap (Dice) while clinical deployment increasingly demands instance-level detection. Small lesions are particularly affected because a near-miss prediction scores the same as a complete miss under instance-matching thresholds. The team also introduces Viola2Plus, a resolution-aware attention architecture for small-lesion segmentation. In an interesting separation of concerns, Viola2Plus left small-lesion Dice unchanged but raised small-lesion detection rate by 3.7% — a gain invisible to voxel-overlap metrics alone. On 5-fold cross-validation across 1,453 training cases, their two-architecture ensemble with VCAP achieves Dice 0.651 and Lesion-F1 0.614, versus 0.644 and 0.573 for the unprocessed single-model baseline. The finding that evaluation unit choice determines what gets optimised has implications beyond stroke: it applies to any segmentation pipeline where detection completeness matters alongside overlap accuracy.