Deep learning-based automatic bleeding recognition during liver resection in laparoscopic hepatectomy

Background: Intraoperative hemorrhage during laparoscopic hepatectomy (LH) is a risk factor for negative postoperative outcomes. Ensuring appropriate hemostasis enhances the safety of surgical procedures. An automatic bleeding recognition system based on deep learning can lead to safer surgeries; however, deep learning models that are useful for detecting and stopping bleeding in LH have not yet been reported. In this study, we aimed to develop a deep learning model to automatically recognize bleeding regions during liver transection in LH.

Methods: In this retrospective feasibility study, bleeding scenes were randomly selected from LH videos, and the videos were divided into frames at 30 frames per second. Bleeding regions within the images were annotated by pixels, and subsequently, all images were assigned to the training, validation, and test datasets to develop the deep learning model. A convolutional neural network algorithm was used to perform semantic segmentation. After training and validation, the model was evaluated using images from the test dataset. Precision, recall, and Dice coefficients served as the evaluation metrics for the model.

Results: In total, 2203 annotated images from 44 LH videos were utilized and divided into 1500, 400, and 303 frames for the training, validation, and test datasets, respectively. The precision, recall, and Dice coefficient values of the model were 0.76, 0.79, and 0.77, respectively.

Conclusions: We developed an automatic bleeding recognition model based on semantic segmentation and verified its performance. The proposed model is potentially useful for intraoperative alerting or evaluating surgical skills in the future.