588 - Semi-Supervised Learning Achieves Fully Supervised Performance for Infarct Segmentation Using 50% Fewer Annotations

Purpose: Expert annotation limits deployment of AI stroke segmentation. We evaluated whether semi‑supervised learning (SSL) can maintain accuracy with half the labeled data and report patient‑level volume agreement.

Materials and Methods: We implemented a Mean‑Teacher nnU‑Net on ISLES2022 MRI (FLAIR/DWI/ADC; 250 cases, 3 centers, 5 scanners).{1,2} Models using 25%, 50%, or 100% labels were compared with a fully supervised nnU‑Net. Data were split 200 train/50 held‑out validation, registered to MNI‑152 with standard augmentations. Primary outcome was Dice on the validation set; paired Wilcoxon tests compared models. We also performed Bland-Altman analysis of predicted vs ground‑truth core volume.

Results: With 50% labels, SSL achieved Dice 0.644±0.243 vs 0.640±0.258 for full supervision (Wilcoxon p=0.664; no significant difference). At the same 50% label fraction, SSL outperformed supervised learning (0.644 vs 0.539; p< 0.01). With 25% labels, SSL (0.612±0.239) exceeded supervised learning with 50% labels (p< 0.01). Bland-Altman on the 50‑case validation set showed a small positive bias of +4.1 mL (95% CI -1.1 to +9.4 mL) with 95% limits of agreement -32.3 to +40.5 mL (each LoA 95% CI ±9.1 mL); dispersion increased for very large cores.

Conclusion: SSL maintained segmentation performance while halving labeled data. Volume agreement demonstrates minimal average bias, with wider errors mainly in very large cores-suggesting limited impact on triage thresholds except for outliers. Halving labels in our 200‑case training set eliminates ~100 manual segmentations saving an estimated 17-32 annotator hours (assuming 10-19 min/case){3,4} and $2800-$5400 (mean US Radiologist wages) {5}, reducing burden without degrading accuracy. Next steps include 5‑fold cross‑validation, an external/site‑held‑out test, and formal equivalence testing (TOST).