Related Experiment Video
Updated: Apr 26, 2026
A Novel Dual-Modal Deep Learning Approach for Real-Time Removal of Hepatic Fluorescence in Indocyanine Green-Guided Laparoscopic Cholecystectomy
Published on: April 17, 2026
Few-shot learning for surgical phase recognition: Performance and generalization in cholecystectomy
Flakë Bajraktari1, Robert Asmußen1, Giuliano A Giacoppo1
1Institute of Medical Device Technology, University of Stuttgart, Pfaffenwaldring 9, 70569, Stuttgart, Germany.
Background And Objective:
The automated recognition of surgical phases in intraoperative videos represents a critical milestone in the digital transformation of surgery. It forms the foundation for surgical assistance systems and enhanced decision support, contributing to increased safety, efficiency, and precision in surgical procedures. Traditional deep learning methods often fall short in this domain due to their dependency on extensive annotated datasets, which are challenging to obtain in medical contexts due to privacy concerns and data scarcity. This study explores the potential of few-shot learning as a paradigm for overcoming data limitations in surgical phase recognition.
Methods:
By leveraging the ability to generalize from minimal examples, a transformer-based few-shot learning (FSL) model for action recognition was adapted to the recognition of surgical phases using the Cholec80 dataset, which consists of videos from cholecystectomy surgeries. The model's performance was evaluated across three experimental splits to assess domain-specific and cross-domain performance: Split 1, where the model was trained on a surgical dataset and tested on Cholec80; Split 2, which introduced variations in surgical environments; and Split 3, where the model was trained on action recognition data and tested on surgical data.
Results:
The model achieved test accuracies of 89.0%, 75.4%, and 49.1% in these splits, respectively. While FSL demonstrates strong applicability to surgical data, domain-specific training remains crucial for optimal performance. Notably, these results were obtained using only a few labeled support examples per phase, illustrating the data efficiency of the approach.
Conclusion:
This study provides an initial foundation for applying few-shot learning to surgical phase recognition and demonstrates its feasibility under low-label and transfer settings. While domain-specific training remains important, the results indicate that FSL is a promising direction for surgical workflow analysis when annotation resources are limited.
