Machine learning-based detection of electrosurgical device-induced bleeding in laparoscopic videos of diaphragmatic hernia repair.

BACKGROUND

Electrosurgical devices provide significant advantages for tissue dissection in laparoscopic procedures. However, achieving optimal hemostasis while minimizing tissue coagulation is challenging. Monitoring device-induced bleeding will provide viable information for surgical skills assessment. We aimed to automatically detect bleeding induced by electrosurgical device use in laparoscopic videos using machine learning.

METHODS

We present a two-step methodology for the automated detection of device-induced bleeding. First, based on the color representation, a random forest classifier (RFC) detects blood pixels in the frames before and after the electrosurgical device activation. Subsequently, a logistic regression (LR) model decides whether bleeding has occurred based on the change in blood pixels. The moments of device activations during surgery can be extracted automatically by a synchronized recording of the laparoscopic video and energy generator data. The RFC and LR were developed on the manual annotation of 34 images and 2678 video fragments from forty-five patients who underwent diaphragmatic hernia repair between May 2023 and October 2024. The performance of the RFC was evaluated by an 80/20 split for training and testing, while a stratified threefold cross-validation assessed the LR performance.

RESULTS

The blood pixel detection showed an accuracy of 94% and a Dice score of 0.472. The classification of automatically extracted video fragments showed that device-induced bleeding can be detected with a 78.2% accuracy, 4.6% precision, 78.1% specificity, and 81.0% sensitivity.

CONCLUSION

The presented work on device-induced bleeding detection is a step toward quantifying the effect of electrosurgery use. We showed a machine learning-based methodology that accurately identifies video fragments of device activations without bleeding but struggles to identify bleeding precisely. Future work should focus on developing device-induced bleeding detection in a larger, more diverse dataset.