Localizing the Center of Surgical Action in Laparoscopic Videos: A Point-Supervised Heatmap Regression Approach

Amir Ebrahimzadeh, Nazila Esmaeili, Philipp Horvath, Michael Ghadimi, Jannis Hagenah

Keywords: Laparoscopic camera navigation, Center of surgical action, Heatmap regression, Point localization, Automated skill assessment

TL;DR: To automate laparoscopic camera centering assessment, this study formalizes center of surgical action localization via heatmap regression, showing a DINOv3-ViT-S model predicts coordinates within 20% of the frame diagonal in 95% of cases.

Registration Requirement: Yes

Abstract: Automated assessment of laparoscopic camera centering requires reliable spatial localization of the center of surgical action (COSA). We formalize this task as a point-localization problem addressed via heatmap-regression. We introduce a dataset of 500 annotated laparoscopic video clips and evaluate three encoder-decoder architectures (UNet-ResNet-34, Segformer-MiT-B1, and a custom DINOv3-ViT-S). DINOv3 achieved the best performance, with approximately 95\% of detections deviating less than 20\% of the frame diagonal from ground truth. This establishes a strong baseline for the COSA localization task and provides an open benchmark for future camera-navigation research.

Reproducibility: https://gitlab.gwdg.de/cds/lcn-cosa-localization

Visa & Travel: Yes

Read CFP & Author Instructions: Yes

Originality Policy: Yes

Single-blind & Not Under Review Elsewhere: Yes

LLM Policy: Yes

Submission Number: 58