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

  • 1

    The proposed Scene Graph-Guided Uncertainty Decomposition (SG-UD) framework improves confidence calibration in surgical visual question answering.

  • 2

    SG-UD decomposes uncertainty into object-level, relation-level, and scene-level components for enhanced interpretability.

  • 3

    Experiments show SG-UD achieves 63.58% accuracy and significantly improves Brier Score and NLL compared to baseline methods.

  • 4

    The framework incorporates Dirichlet-based calibration, which is critical for enhancing probabilistic quality in predictions.

  • 5

    Ablation studies indicate that uncertainty decomposition contributes most to answer accuracy, improving it by 0.70%.

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

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

Computational modelling for personalized transcatheter aortic valve replacement planning: a systematic review of complications and decision support

Development and validation of a predictive model for postoperative delirium in patients undergoing cardiac surgery