Objective To investigate the application value of major anatomical structure recognition model of minimally invasive liver resection based on deep learning.

Methods The retrospective and descriptive study was conducted. The 31 surgical videos of laparoscopic left lateral sectionectomy performed in Zhujiang Hospital of Southern Medical University from January 2019 to April 2023 were collected. Video clips containing the surgical procedure of left lateral lobe liver pedicle and left hepatic vein were screened by 2 liver surgeons. After quality control, screening and frame extraction, the major anatomical structures on the images of these clips were annotated. After pre‑processing, these images were transported to the DeepLab v3+neural network framework for model training. Observation indicators: (1) video annotation and classification; (2) results of arti-ficial intelligence anatomical recognition model testing. Measurement data with normal distribution were represented as Mean±SD, and count data were described as absolute numbers.

Results (1) Video annotation and classification. A total of 4 130 frames of images were annotated in the 31 surgical videos, including 2 083 frames of annotated images for the left lateral lobe liver pedicle, 1 578 frames of annotated images for the left hepatic vein and 469 frames of annotated images for both the left lateral lobe liver pedicle and left hepatic vein. (2) Results of artificial intelligence anatomical recognition model testing. In four application scenarios (clean scene, bloodstain scene, partially obstruction by instrument scene, and small exposed area scene), the model was able to successfully recognize the left lateral lobe liver pedicle and left hepatic vein, with a recognition speed for anatomical markers >13 frames/s. When performing anatomical recognition on images with only the left lateral lobe liver pedicle, the Dice coefficient, intersection over union, accuracy, sensitivity and specificity of the model were 0.710±0.110, 0.560±0.120, 0.980±0.010, 0.640±0.030, and 0.980±0.010, respectively. The above indicators of the model were 0.670±0.180, 0.530±0.200, 0.980±0.010, 0.600±0.040, and 0.990±0.010 when performing anatomical recognition on images with only the left hepatic vein, and 0.580±0.180, 0.430±0.190, 0.980±0.010, 0.580±0.020, and 0.990±0.010 when per-forming anatomical recognition on images with both the left lateral lobe liver pedicle and left hepatic vein.

Conclusion The major anatomical structure recognition model of minimally invasive liver resection based on deep learning can be applied in identifying liver pedicle and hepatic vein.