EAES Academy

Background:
Because thoracic esophageal cancer has a high metastatic rate to the upper mediastinal lymph nodes, it is essential to perform complete lymphadenectomy around Recurrent laryngeal nerve (RLN) without complication, especially RLN paralysis (RLNP). It is important for surgeons to quickly and accurately identify the RLN during surgery, therefore intraoperative image navigation may be useful. The purpose of this study was to develop a deep learning model for RLN identification by semantic segmentation based on thoracoscopic videos, and clarify the accuracies of automatic RLN nerve identification.
Methods:

Semantic segmentation of the RLN area was performed using a convolutional neural network(CNN)-based approach. DeepLab v3 plus was utilized as the CNN model for the semantic segmentation task. The Dice coefficient (DC) was utilized as an evaluation metric for the proposed model. We conducted the comparative study between Artificial Intelligence (AI) and surgeons (expert surgeons/general surgeon) to validate the accuracy of RLN identification.
Results:

Three thousand RLN images were randomly extracted from 20 videos of thoracoscopic esophagectomy, and the RLN area was manually annotated on each image. The average DC for RLN in this AI model was 0.58. In the comparative study, the average DC were 0.58 for AI, 0.62 for expert surgeons and 0.47 for general surgeons. There was no significant difference in dice score for RLN between AI and expert surgeons(p=0.38), but dice score in this AI model was significantly higher than it in general surgeons(p<0.05).
Conclusions:

This result suggest that AI may be useful in identifying the RLN during thoracoscopic esophagectomy. If the accuracy of AI is improved in the future, it may help safe surgery for thoracoscopic esophagectomy. Currently, we perform intraoperative real-time RLN identification using AI, and work to improve accuracy for clinical application(Fig. 1, 2).