Scene graph-guided uncertainty decomposition improves confidence calibration in surgical visual question answering - Summary - MDSpire

Scene graph-guided uncertainty decomposition improves confidence calibration in surgical visual question answering

  • By

  • Junzhuo Song

  • Yaoxue Xu

  • Zihan Zhu

  • Junshuang Zhou

  • Yuhaohang He

  • Xirui Chen

  • Yunsen Liang

  • Wei Chen

  • Qiurui Liu

  • Jun Li

  • June 5, 2026

  • 0 min

Share

Objective:

To improve confidence estimation in surgical visual question answering by decomposing uncertainty into hierarchical components.

Key Findings:
  • Achieved an overall accuracy of 63.58% on the SSG-QA benchmark.
  • Expected Calibration Error of 16.97% and Risk-Coverage AUC of 0.0861.
  • Significant improvements in Brier Score and NLL over baseline methods.
  • Largest accuracy gain observed for relation-type questions (+0.92% over MCAN).
  • Uncertainty decomposition contributed most to answer accuracy (+0.70%).
Interpretation:

Limitations:
  • The study may not address all potential sources of uncertainty in surgical VQA.
  • Further exploration of ethical considerations and future directions is needed.
Conclusion:

Original Source(s)

Scene graph-guided uncertainty decomposition improves confidence calibration in surgical visual question answering

  • by Junzhuo Song, Yaoxue Xu, Zihan Zhu, Junshuang Zhou, Yuhaohang He, Xirui Chen, Yunsen Liang, Wei Chen, Qiurui Liu, Jun Li

  • June 5, 2026

Related Content

Application of a self-developed femoral artery compression hemostasis device in proximal femoral nail anti-rotation surgery for intertrochanteric fractures: a case report

Development and validation of a clinical prediction model for postoperative atrial fibrillation after lung cancer surgery: a machine-learning–based study

Biportal endoscopic spinal surgery for thoracic ossification of the ligamentum flavum: a study of different classification types and surgical outcomes