Accuracy of Artificial Intelligence-based Surgical Support Systems in Recognizing Thoracic Nerves

J. Ichinose¹, N. Kobayashi², K. Fukada², K. Kanno², A. Suzuki¹, Y. Matsuura¹, M. Nakao¹, S. Okumura¹, M. Mun^1
1Cancer Institute Hospital of JFCR, Koto-ku, Tokyo ²Anaut Inc., Minato-ku, Tokyo

Purpose: We have developed a surgical support system that analyzes the surgical field using artificial intelligence (AI) to visualize the important microanatomy such as the dissection layer, nerves, blood vessels, ureters, and pancreas.1,2 Herein, we evaluated the accuracy of the system in recognizing thoracic nerves during lung cancer surgery.

Methods: This prospective, observational study included patients who underwent thoracoscopic lobectomy or segmentectomy with mediastinal lymphadenectomy for lung cancer between June 2022 and April 2023. The surgery was performed using four ports with confronting monitor setting. The AI system consists of a workstation and display monitor, which can be connected to a surgical endoscope system using a single cable. Recognition models were created by deep learning using images precisely annotated by board-certified surgeons for each anatomical structure. The Dice index and Jaccard index evaluated the degree of agreement between the AI recognition and annotated areas.
AI recognition result images were presented in real time during thoracoscopic surgery. Additionally, the differences in time lag, image quality, and smoothness of movement between the AI system and surgical monitor were evaluated. Four general thoracic surgeons evaluated the accuracy of nerve recognition for sensitivity and specificity and ratings were made on a five-point scale.

Results: More than 6,000 annotated images were used to create the recognition models, and 682 annotated images were used for nerves’ recognition. The computational evaluation of the created AI recognition model of the thoracic nerves was relatively favorable for recognizing numerous thin structures, with a Dice index of 0.56 and Jaccard index of 0.39. The surgical support system was used in 7 left lung surgery and 3 right lung surgery cases and presented thoracic nerves in all. It clearly presented not only the vagus, phrenic, and left recurrent nerves, which are important for mediastinal lymph node dissection, but also their minor branches. A difference in the smoothness of movement (3.2 ± 0.4 points) was noticeable although almost no difference in time lag (4.9 ± 0.3 points) or image quality (4.6 ± 0.5 points) between the AI system and surgical monitor was perceived. The accuracy of thoracic nerve recognition was satisfactory with a sensitivity score of 4.5 ± 0.4 and specificity score of 4.0 ± 0.9. There was a tendency to misidentify parts of the pleura and bronchial wall as nerves.

Conclusion: The AI surgical support system demonstrated a satisfactory accuracy in recognizing the thoracic nerves for the expert surgeons. The recognition results could be presented during surgery, with no time lag or image quality degradation. This system may allow any surgeon to perform safe surgery in the future.

Identify the source of the funding for this research project: None.