Chen Wei, Fukuda Shohei, Yoshida Soichiro, Kobayashi Nao, Fukada Kyohei, Fukunishi Munenori, Otani Yuhi, Matsumoto Shunya, Kobayashi Masaki, Nakamura Yuki, Fan Bo, Ishikawa Yudai, Fukushima Hiroshi, Fu Guangqing, Waseda Yuma, Tanaka Hajime, Fujii Yasuhisa
Department of Urology, Institute of Science Tokyo, Tokyo, Japan.
Department of Urology, Zigong Fourth People's Hospital, Sichuan, China.
J Robot Surg. 2025 Apr 30;19(1):188. doi: 10.1007/s11701-025-02340-2.
Artificial intelligence (AI)-driven intraoperative navigation in urological surgery can enhance surgical precision through real-time structure identification and tracking. This study describes a novel AI solution that enables real-time fluorescence-like navigation (FLN) for robot-assisted radical cystectomy (RARC) with an initial focus on ureter, potentially enhancing outcomes and training efficacy. We established a new AI model using convolutional neural network (CNN) to achieve real-time intraoperative ureter recognition using 730 images from 17 RARC cases. Quantitative and qualitative analyses were performed for each procedure (phases I-V: identifying, exposing, elevating, retracting, distal separation). For quantitative evaluation, we calculated precision, recall, intersection over union (IoU), and Dice coefficients by comparing AI-inferred images with surgeons' annotations on 41 test images. In addition, 18 surgeons participated in a qualitative assessment, answering questions on identification, misidentification, and clinical utility. The CNN-based prediction model has been successfully established and validated. The AI model achieved a Dice score of 0.71 (phases I-V, respectively: 0.75/0.69/0.71/0.69/0.72), IoU of 0.55 (0.60/0.53/0.55/0.52/0.57), recall of 0.90 (0.91/0.92/0.92/0.89/0.89), precision of 0.60 (0.67/0.55/0.58/0.61/0.64), and accuracy of 0.99 (1.00/0.99/0.99/0.99/0.99). For identification performance, the AI system scored an average of 4.74 on a scale of 0 to 5 (4.94/4.61/4.94/4.89/4.33), indicating that most images achieved > 80% recognition accuracy. The average score for misidentification of other tissues as ureter was low at 0.60 on a scale of 0 to 5 (phases I-V: 0.11/0.56/0.39/0.44/1.50). In the clinical utility assessment, 62.22% (60.00%-64.44%) of the AI-inferred images were correctly distinguished from the ground truth. Our AI model reliably annotated the ureter in real-time during RARC, achieving high accuracy and acceptable precision. This technology has the potential to reduce the risk of ureter misrecognition by surgeons, thereby enhancing surgical accuracy and safety.
人工智能（AI）驱动的泌尿外科手术术中导航可通过实时结构识别和跟踪提高手术精度。本研究描述了一种新型AI解决方案，该方案可为机器人辅助根治性膀胱切除术（RARC）实现实时类荧光导航（FLN），最初重点关注输尿管，有望提高手术效果和训练效率。我们使用卷积神经网络（CNN）建立了一个新的AI模型，利用17例RARC病例的730张图像实现术中输尿管的实时识别。对每个手术步骤（I - V期：识别、暴露、提起、牵拉、远端分离）进行了定量和定性分析。为进行定量评估，我们通过将AI推断图像与41张测试图像上外科医生的标注进行比较，计算了精确率、召回率、交并比（IoU）和Dice系数。此外，18名外科医生参与了定性评估，回答了关于识别、误识别和临床实用性的问题。基于CNN的预测模型已成功建立并得到验证。AI模型的Dice评分为0.71（I - V期分别为：0.75/0.69/0.71/0.69/0.72），IoU为0.55（0.60/0.53/0.55/0.52/0.57），召回率为0.90（0.91/0.92/0.92/0.89/0.89），精确率为0.60（0.67/0.55/0.58/0.61/0.64），准确率为0.99（1.00/0.99/0.99/0.99/0.99）。对于识别性能，AI系统在0至5分的评分中平均得分为4.74（4.94/4.61/4.94/4.89/4.33），表明大多数图像的识别准确率超过80%。将其他组织误识别为输尿管的平均得分在0至5分的评分中较低，为0.60（I - V期：0.11/0.56/0.39/0.44/1.50）。在临床实用性评估中，62.22%（60.00% - 64.44%）的AI推断图像与真实情况能够正确区分。我们的AI模型在RARC过程中能够可靠地实时标注输尿管，实现了高精度和可接受的精确率。这项技术有可能降低外科医生误认输尿管的风险，从而提高手术的准确性和安全性。
