Abstract
Transformer-based methods have improved surgical phase recognition, but their large model size and high computational cost make them less practical for long surgical videos and resource-constrained set tings. This paper investigates whether a lightweight student model can preserve temporal modeling ability while substantially reducing the parameter count and per-clip computation of SurgFormer. We propose a SF-Small+BiGRU framework that combines compact Transformer-based clip encoding with recurrent cross-clip temporal modeling. The resulting model reduces the parameter count from 177.8M to 31.9M and the per clip computation from 893.28 to 150.24 GFLOPs. On Cholec80, the cross entropy (CE)-trained model achieved 0.8956 accuracy and 0.7143 Jaccard, with a Jaccard decrease of 0.0175 compared with the SurgFormer teacher despite the substantial reduction in model size. On AutoLaparo, GRU-level distillation further improved Jaccard from 0.4997 to 0.5257. These results suggest that, under the evaluated settings, recurrent cross clip temporal modeling contributes more consistently to the lightweight model than knowledge distillation, whose effects are dataset- and level dependent.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/COLAS_006.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=Jkvs5xQQWx
BibTex
@InProceedings{LiLun_Efficient_MICCAISAT2026,
author = { Li, Luning AND Sanmoto, Yohei AND Goto, Yudai AND Zhu, Xuhan AND Jin, Qun AND Peng, Bo AND Zhou, Zhiyong AND Dai, YaKang AND Xu, Rong AND Zhu, Xin},
title = { { Efficient Surgical Phase Recognition via Compact Transformer Encoding and Recurrent Temporal Modeling } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS pending},
month = {pending},
page = {pending}
}