An Automated Machine Learning Framework for Surgical Suturing Action Detection under Class Imbalance

TL;DR

This paper addresses real-time detection of laparoscopic surgical actions under class imbalance with a need for interpretability and reliable feedback. It proposes a rapid-deploy AutoML framework that frames surgical action recognition as a CASH problem, uses resampling and class weighting to handle imbalance, and builds robust ensembles via meta-learning warm starts. The approach demonstrates high accuracy and stability on suturing tasks across experienced and novice surgeons, outperforming several deep-learning baselines while preserving real-time inference capabilities. The work has practical implications for automated, intelligent feedback in laparoscopic training and suggests avenues for dataset expansion and integration with newer detection methods.

Abstract

In laparoscopy surgical training and evaluation, real-time detection of surgical actions with interpretable outputs is crucial for automated and real-time instructional feedback and skill development. Such capability would enable development of machine guided training systems. This paper presents a rapid deployment approach utilizing automated machine learning methods, based on surgical action data collected from both experienced and trainee surgeons. The proposed approach effectively tackles the challenge of highly imbalanced class distributions, ensuring robust predictions across varying skill levels of surgeons. Additionally, our method partially incorporates model transparency, addressing the reliability requirements in medical applications. Compared to deep learning approaches, traditional machine learning models not only facilitate efficient rapid deployment but also offer significant advantages in interpretability. Through experiments, this study demonstrates the potential of this approach to provide quick, reliable and effective real-time detection in surgical training environments

Figures (5)

• Figure 1: Overall AutoML workflow including meta learning warmstart for bayesian optimization efficient model selection and ensemble building for laparoscopy surgical suturing action detection.
• Figure 2: Class distribution across action categories shows a noticeable imbalance, with certain categories, such as Double throw and Set needle, containing significantly more samples compared to others like Grasp needle and Needle exits. This imbalance may lead to biased model performance, favoring well-represented classes while potentially underperforming on underrepresented ones.
• Figure 3: Model accuracy comparison, illustrating the impact of automated optimization versus conventional deep learning approaches.
• Figure 4: Analysis of prediction speed and accuracy relative to model size in ensemble architectures
• Figure 5: Experienced surgeons' trajectories show consistent, controlled, and concentrated movements within a defined spatial range, reflecting precision and coordination. In contrast, novice surgeons' trajectories display irregular and dispersed patterns, suggesting a lack of control and coordination. The trajectories of novice surgeons also demonstrate frequent abrupt deviations and scattered points, indicating variability and more erratic hand movements. These inconsistencies and outliers in novice trajectories may introduce noise to the data, potentially impacting model training by skewing predictions toward these deviations.