Smarter Self-Distillation: Optimizing the Teacher for Surgical Video Applications

This repository provides reproduction code for the following research papers:

• Self-Distillation for Surgical Action Recognition
  Presented at MICCAI 2023.

  • To reproduce the results from this paper, please refer to the repository tag: miccai23.
  • Additional documentation is available in readme_miccai23.md.

• Smarter Self-Distillation: Optimizing the Teacher for Surgical Video Applications
  Published at MICCAI 2025. See the preprint for details.

---

What's New in the MICCAI 2025 Release

The MICCAI 2025 implementation improves self-distillation performance by optimizing the teacher model, overcoming the limitations of our previous self-distillation method (MICCAI 2023 work), which relied on a sub-optimal teacher without teacher selection.

This release introduces two key improvements:

• Teacher Selection – automatically identifies the most effective teacher model (checkpoint) for distillation.
• Multi-Teacher Ensemble – leverages multiple teachers to enhance the quality of soft labels.

---

Code Availability

The code for the MICCAI 2025 paper will be released shortly after the conference.

---

Reproducing MICCAI 2023 Results

To reproduce the experiments from Self-Distillation for Surgical Action Recognition (MICCAI 2023):

• Switch to the repository tag: miccai23
• Refer to the documentation in readme_miccai23.md

---