Xu Xiayue, Yun Boxiang, Zhao Yumin, Jin Ling, Zong Yanning, Yu Guanzhen, Zhao Chuanliang, Fan Kai, Zhang Xiaolin, Tan Shiwang, Zhang Zimu, Wang Yan, Li Qingli, Yu Shaoqing
Department of Otolaryngology and Neck Surgery, Tongji Hospital, School of Medicine, Tongji University, Shanghai 200065, China.
Department of Allergy, Tongji Hospital, School of Medicine, Tongji University, Shanghai 200065, China.
Bioengineering (Basel). 2024 Dec 25;12(1):10. doi: 10.3390/bioengineering12010010.
We aim to construct an artificial intelligence (AI)-assisted nasal endoscopy diagnostic system capable of preliminary differentiation and identification of nasal neoplasia properties, as well as intraoperative tracking, providing an important basis for nasal endoscopic surgery.
We retrospectively analyzed 1050 video data of nasal endoscopic surgeries involving four types of nasal neoplasms. Using Deep Snake, U-Net, and Att-Res2-UNet, we developed a nasal neoplastic detection network based on endoscopic images. After deep learning, the optimal network was selected as the initialization model and trained to optimize the SiamMask online tracking algorithm.
The Att-Res2-UNet network demonstrated the highest accuracy and precision, with the most accurate recognition results. The overall accuracy of the model established by us achieved an overall accuracy similar to that of residents (0.9707 ± 0.00984), while slightly lower than that of rhinologists (0.9790 ± 0.00348). SiamMask's segmentation range was consistent with rhinologists, with a 99% compliance rate and a neoplasm probability value ≥ 0.5.
This study successfully established an AI-assisted nasal endoscopic diagnostic system that can preliminarily identify nasal neoplasms from endoscopic images and automatically track them in real time during surgery, enhancing the efficiency of endoscopic diagnosis and surgery.
构建一种人工智能（AI）辅助鼻内镜诊断系统，能够对鼻腔肿瘤性质进行初步鉴别与识别，并在术中进行跟踪，为鼻内镜手术提供重要依据。
回顾性分析1050例涉及四种鼻腔肿瘤的鼻内镜手术视频数据。利用深度蛇形模型、U-Net和Att-Res2-UNet，我们基于内镜图像开发了一种鼻腔肿瘤检测网络。经过深度学习后，选择最优网络作为初始化模型，并对SiamMask在线跟踪算法进行训练优化。
Att-Res2-UNet网络表现出最高的准确率和精确率，识别结果最为准确。我们建立的模型总体准确率与住院医师相似（0.9707±0.00984），但略低于鼻科医生（0.9790±0.00348）。SiamMask的分割范围与鼻科医生一致，符合率达99%，肿瘤概率值≥0.5。
本研究成功建立了一种AI辅助鼻内镜诊断系统，该系统能够从内镜图像中初步识别鼻腔肿瘤，并在手术过程中实时自动跟踪，提高了内镜诊断和手术的效率。
