Department of Urology, Severance Hospital, Urological Science Institute, Yonsei University College of Medicine, Seoul 03722, Republic of Korea.
Department of Urology, Yeungnam University College of Medicine, Daegu 42415, Republic of Korea.
Medicina (Kaunas). 2023 Jul 30;59(8):1400. doi: 10.3390/medicina59081400.
Analysis of urine stone composition is one of the most important factors in urolithiasis treatment. This study investigated whether a convolutional neural network (CNN) can show decent results in predicting urinary stone composition even in single-use flexible ureterorenoscopic (fURS) images with relatively low resolution. This study retrospectively used surgical images from fURS lithotripsy performed by a single surgeon between January 2018 and December 2021. The ureterorenoscope was a single-use flexible ureteroscope (LithoVue, Boston Scientific). Among the images taken during surgery, a single image satisfying the inclusion and exclusion criteria was selected for each stone. Cases were divided into two groups according to whether they contained any calcium oxalate (the Calcium group) or none (the Non-calcium group). From 506 total cases, 207 stone surface images were finally included in the study. In the CNN model, the transfer learning method using Resnet-18 as a pre-trained model was used, and only endoscopic digital images and stone classification data were input to achieve minimally supervised learning. There were 175 cases in the Calcium group and 32 in the Non-calcium group. After training and validation, the model was tested using the test set, and the total accuracy was 81.8%. Recall and precision of the test results were 88.2% and 88.2% in the Calcium group and 60.0% and 60.0% in the Non-calcium group, respectively. The area under the receiver operating characteristic curve of the model, which represents its classification performance, was 0.82. Single-use flexible ureteroscopes have financial benefits but low vision quality compared with reusable digital flexible ureteroscopes. As far as we know, this is the first artificial intelligence study using single-use fURS images. It is meaningful that the CNN performed well even under these difficult conditions because these results can further expand the possibilities of its use.
尿石成分分析是尿石症治疗中最重要的因素之一。本研究旨在探讨卷积神经网络（CNN）是否即使在单次使用的软性输尿管镜（fURS）图像分辨率较低的情况下，也能在预测尿石成分方面取得较好的效果。本研究回顾性地使用了 2018 年 1 月至 2021 年 12 月期间由同一位外科医生进行的 fURS 碎石术的手术图像。输尿管镜为一次性使用的软性输尿管镜（LithoVue，Boston Scientific）。在手术过程中拍摄的图像中，每个结石选择符合纳入和排除标准的单个图像。根据结石是否含有任何草酸钙（钙组）或不含有任何草酸钙（非钙组）将病例分为两组。在 506 例总病例中，最终纳入 207 例结石表面图像。在 CNN 模型中，使用 Resnet-18 作为预训练模型的迁移学习方法，仅输入内镜数字图像和结石分类数据，实现最小监督学习。钙组有 175 例，非钙组有 32 例。经过训练和验证后，使用测试集对模型进行测试，总准确率为 81.8%。钙组的测试结果的召回率和准确率分别为 88.2%和 88.2%，非钙组分别为 60.0%和 60.0%。模型的分类性能代表其接收者操作特征曲线下面积为 0.82。与可重复使用的数字软性输尿管镜相比，一次性使用的软性输尿管镜具有经济优势，但视觉质量较低。据我们所知，这是首次使用单次使用的 fURS 图像进行人工智能研究。令人感到有意义的是，即使在这些困难的条件下，CNN 也能表现良好，因为这些结果可以进一步扩大其应用的可能性。
