EUVIMEDEUVIMED entdecken
Crossref

SAGE: Self-Evolving Acquisition Through Guided Escalation for Subtle Anomaly Segmentation

Renjith Prasad Kaippilly Mana, Rishabh Sharma, Andrew Shao, Annmary Justine Koomthanam, Shreyas Kulkarni, Suparna Bhattacharya, Martin Foltin, Amit Sheth, David Orozco, Matthew Quinn, Brian Sammuli

Proceedings of the AAAI Symposium Series · 06.11.2026 · 10 (1) · 626-632

DOI: 10.1609/aaaiss.v10i1.43465 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Proceedings of the AAAI Symposium Series, online erschienen am 06.11.2026. Verfasst von Renjith Prasad Kaippilly Mana und 10 weiteren Autor:innen.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Detecting subtle, weak-signal structures in scientific imaging, such as low-contrast irregularities in CT scans, faint abnormalities in microscopy, or rare targets in large survey data, is often limited less by model capacity than by the cost of expert annotation. The structures of interest are small, weakly visible, and frequently predicted as background, so standard active learning methods based on uncertainty or diversity can overlook precisely the regions most useful for improving the next model. We present SAGE, a self-evolving active learning framework for region-level subtle anomaly segmentation under sparse supervision. SAGE conditions acquisition on the current segmentation state, uses a fast vision--language policy to rank candidate regions, andselectively escalates ambiguous or redundancy-sensitive decisions to a fixed deliberative planner. After annotation and retraining,reward-set improvements are converted into pairwise preferences that update the fast policy through outcome-validated preference learning. On CT-Defect and two public subtle-defect benchmarks, MVTec AD and VisA,SAGE improves subtle-defect Dice by 6.6--7.2 points over the strongest learned region-selection baseline and surfaces confirmed targets earlierunder a fixed review budget. These results show that acquisition policies can be improved from measured downstream segmentation outcomes, providing a data-efficient mechanism for focusing scarce expertsupervision on subtle visual evidence.

Zitieren

APA

Mana, R. P. K., Sharma, R., Shao, A., Koomthanam, A. J., Kulkarni, S., Bhattacharya, S., Foltin, M., Sheth, A., Orozco, D., Quinn, M., & Sammuli, B. (2026). SAGE: Self-Evolving Acquisition Through Guided Escalation for Subtle Anomaly Segmentation. Proceedings of the AAAI Symposium Series, 10(1), 626-632. https://doi.org/10.1609/aaaiss.v10i1.43465

Vancouver

Mana RPK, Sharma R, Shao A, Koomthanam AJ, Kulkarni S, Bhattacharya S, et al. SAGE: Self-Evolving Acquisition Through Guided Escalation for Subtle Anomaly Segmentation. Proceedings of the AAAI Symposium Series. 2026;10(1):626-632. doi:10.1609/aaaiss.v10i1.43465

BibTeX
@article{Mana2026,
  author = {Mana, Renjith Prasad Kaippilly and Sharma, Rishabh and Shao, Andrew and Koomthanam, Annmary Justine and Kulkarni, Shreyas and Bhattacharya, Suparna and Foltin, Martin and Sheth, Amit and Orozco, David and Quinn, Matthew and Sammuli, Brian},
  title = {SAGE: Self-Evolving Acquisition Through Guided Escalation for Subtle Anomaly Segmentation},
  journal = {Proceedings of the AAAI Symposium Series},
  year = {2026},
  volume = {10},
  number = {1},
  pages = {626--632},
  doi = {10.1609/aaaiss.v10i1.43465}
}
Kostenlos starten →