Nature Reviews Urology | 2021
The dawning of the age of artificial intelligence in urology
Abstract
0123456789();: Nature reviews | Urology The utility of artificial intelligence (AI) in urology has been shown in two studies published in BJU International. An AI-based algorithm for improving bladder cytology assessment in bladder cancer was developed and a deep learning approach was used to predict lymph node metastasis directly from the primary tumour in prostate cancer. These studies show the potential use of AI in urological diseases. First, the results of the first phase of the VISIOCYT1 clinical trial have been reported. VisioCyt, a digital medical device that can detect tumour cell aspect in voided urine samples, was used and an algorithm developed for assessing bladder cytology slides. In total, 598 individuals were included, 449 with bladder tumours and 149 healthy participants. To train the algorithm, slides were annotated locally and globally. In the local annotation, cells were labelled according to five categories: urothelial cells; squamous cells; inflammatory cells; artefacts; and cell clusters. Then urothelial cells were analysed and marked. For the global annotation, slides were labelled based on the results of the gold-standard examination — a negative slide had negative cytology and negative endoscopy, and a positive slide had histology confirming urothelial neoplasia. The algorithm has five steps: object detection; object classification; local feature computation; global feature computation; and slide classification, in which the slide is scored between 0 and 1 and a value of >0.55 equates to the slide being labelled as positive for bladder cancer. The algorithm was evaluated using the leave one out crossvalidation approach, and the results of the VisioCyt test were compared with voided urine cytology performed by experienced uropathologists. The overall sensitivity, sensitivity for high-grade tumours and sensitivity for lowgrade tumours were all improved using the VisioCyt test compared with cytology (84.9% versus 43%, 92.6% versus 61.1% and 77% versus 26.3%, respectively). Second, a digital biomarker based on analysis of primary prostate tumour tissue by a convolutional neural network (CNN) to predict lymph node metastasis was developed. Stained slides of primary prostate tumour tissue were selected from 218 patients (102 N+ and 116 N0) who were matched for Gleason score, tumour size, venous invasion, perineural invasion and age. The slides were digitized into whole-slide images (WSIs), annotated to exclude the background and artefacts, and tessellated into square patches. In total, 1,000 patches per WSI were used for training and testing. For CNN training, 118 WSIs (60 N+ and 58 N0) were used and were split into a training set (83 WSIs) and a validation set (35 WSIs). A pretrained xse_ResNext34 CNN was used to predict lymph node metastasis. For testing, 100 WSIs were used (42 N+ and 58 N0). Overall, 10 models were trained using the same data. The CNN-based image analysis achieved a mean balanced accuracy of 61.37% for predicting lymph node metastasis. Mean sensitivity and specificity were 53.09% and 69.65%, respectively. The results of these two studies show that use of AI in urology has the potential to aid clinicians in diagnosing disease and to provide relevant prognostic information that could guide patient care.