Development and validation of a deep-learning based assistance system for enhancing laparoscopic control level

Inter-operator variations in the level of intraoperative laparoscope control by surgeons influence surgical outcomes. We aimed to construct a laparoscopic surgery quantification system (LSQS) for real-time evaluation of the surgeon's laparoscope control to improve intraoperative manipulation of the laparoscope.

Using 1888 images from 80 laparoscopic videos for training, the U-Net, PSPNet, LinkNet, and DeepLabv3+ models were used to segment surgical instruments. The percentage of the instruments in central area was defined as the new indicator and the threshold was determined from 20 laparoscopic videos. The differences between expert and non-expert laparoscopic operators before and after LSQS were compared.

Among the three segmentation models (U-Net, PSPNet, and LinkNet), the PSPNet model had the highest index (precision 0.9135; F1 score 0.9058; mIoU 0.8280). The validation experiment showed that LSQS could help non-expert users to more easily achieve expert-level control of the laparoscope.

Deep-learning technology successfully fed back real-time intraoperative information on level of laparoscope control and may facilitate better visualisation of the surgical field.