Abstract
Endovascular videos are challenging: guidewire motions are subtle, informative segments are sparse, and global anatomy must be balanced with local instrument detail. We present CathAgent, a lightweight video framework tailored to these conditions. CathAgent combines three innovations: (1) second-order motion features with adaptive motion fusion, which capture fine-grained dynamics while controlling noisy motion evidence; (2) adaptive temporal weighting, which emphasizes informative segments; and (3) an optional bi-stream input that combines full fluoroscopic frames with guidewire-tip ROIs generated by a YOLOv12 detector. On the CathAction phantom and animal data, the bi-stream model achieves 78.87% accuracy and a macro-F1 score of 0.7496, exceeding the reported scores of the evaluated baselines under the same experimental protocol. The single-stream model retains 76.97% accuracy and runs at 401 frames/s. CathAgent therefore provides an explicit accuracy–speed trade-off for endovascular video analysis. The code and model will be released.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/CREATE_014.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to Open Review
Open Review Page: https://openreview.net/forum?id=Mzv8lGhkY2
BibTex
@InProceedings{CaoTia_CathAgent_MICCAISAT2026,
author = { Cao, Tianyang AND Lin, Kevin Qinghong AND Zhang, Shijie AND Feng, Yuqing AND Nguyen, Anh AND Torr, Philip AND Elson, Daniel S. AND Huang, Baoru},
title = { { CathAgent: A Lightweight Video Assistant for Endovascular Action Recognition } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17275},
month = {pending},
page = {pending}
}