Fernandez Nicolas, Lorenzo Armando J, Rickard Mandy, Chua Michael, Pippi-Salle Joao L, Perez Jaime, Braga Luis H, Matava Clyde
Division of Urology, Seattle Children's Hospital, University of Washington, Seattle, USA.
Department of Surgery, Division of Urology, Hospital for Sick Children, University of Toronto, Canada.
Urology. 2021 Jan;147:264-269. doi: 10.1016/j.urology.2020.09.019. Epub 2020 Sep 26.
To improve hypospadias classification system, we hereby, show the use of machine learning/image recognition to increase objectivity of hypospadias recognition and classification. Hypospadias anatomical variables such as meatal location, quality of urethral plate, glans size, and ventral curvature have been identified as predictors for postoperative outcomes but there is still significant subjectivity between evaluators.
A hypospadias image database with 1169 anonymized images (837 distal and 332 proximal) was used. Images were standardized (ventral aspect of the penis including the glans, shaft, and scrotum) and classified into distal or proximal and uploaded for training with TensorFlow. Data from the training were outputted to TensorBoard, to assess for the loss function. The model was then run on a set of 29 "Test" images randomly selected. Same set of images were distributed among expert clinicians in pediatric urology. Inter- and intrarater analyses were performed using Fleiss Kappa statistical analysis using the same 29 images shown to the algorithm.
After training with 627 images, detection accuracy was 60%. With1169 images, accuracy increased to 90%. Inter-rater analysis among expert pediatric urologists was k= 0.86 and intrarater 0.74. Image recognition model emulates the almost perfect inter-rater agreement between experts.
Our model emulates expert human classification of patients with distal/proximal hypospadias. Future applicability will be on standardizing the use of these technologies and their clinical applicability. The ability of using variables different than only anatomical will feed deep learning algorithms and possibly better assessments and predictions for surgical outcomes.
为改进尿道下裂分类系统，我们在此展示如何使用机器学习/图像识别技术来提高尿道下裂识别与分类的客观性。尿道下裂的解剖学变量，如尿道口位置、尿道板质量、龟头大小和腹侧弯曲，已被确定为术后结果的预测指标，但评估者之间仍存在显著的主观性。
使用了一个包含1169张匿名图像（837例远端型和332例近端型）的尿道下裂图像数据库。图像进行了标准化处理（阴茎腹侧包括龟头、阴茎体和阴囊），并分为远端型或近端型，上传至TensorFlow进行训练。训练数据输出到TensorBoard，以评估损失函数。然后该模型在随机选择的一组29张“测试”图像上运行。同一组图像分发给小儿泌尿外科的专家临床医生。使用与算法展示的相同的29张图像，采用Fleiss Kappa统计分析进行评分者间和评分者内分析。
用627张图像训练后，检测准确率为60%。使用1169张图像时，准确率提高到90%。小儿泌尿外科专家之间的评分者间分析的kappa值为0.86，评分者内为0.74。图像识别模型模拟了专家之间几乎完美的评分者间一致性。
我们的模型模拟了专家对远端/近端尿道下裂患者的分类。未来的适用性将在于规范这些技术的使用及其临床适用性。使用不仅仅是解剖学变量的能力将为深度学习算法提供支持，并可能对手术结果进行更好的评估和预测。
