Medical AI research and development is primarily focused on diagnosis, prediction, treatment, and management as an aid to patient care. AI is primarily being utilized in personal healthcare and diagnostic imaging. In the field of urology, a lot of investment is being made in the development of urination monitoring systems for personal healthcare and diagnostic imaging solutions for ureteral strictures and urinary stones. There are also cases of AI technology being applied to neurogenic bladders. In this paper, we will cover some examples in each of these fields and discuss the prospects for AI applications.

AI research is being conducted on various urological diseases, including prostatic hyperplasia, ureteral stricture, urinary stones [2-10], and urinary tract infections. AI is mainly used as an aid for clinicians to diagnose diseases, and it is essential to verify the effectiveness of applying AI technology through many clinical trials. As a result, we were able to identify fields and technologies with high applicability in the medical field, and in this paper, we introduced some of the major cases by technology field.

First, researchers at Gachon University Gil Hospital developed an assistive technology that can help diagnose urinary diseases by analyzing urination posture, or movement signals, collected through gyro and accelerometer sensors [11]. First, the research team developed a urination diary app that supports the automatic creation of urination diaries for patients with urinary diseases. The urination of male users was categorized into 3 periods: before, during, and after urination, so that the frequency, duration, and volume of the user’s urination could be automatically analyzed through learning and recognition. The technical basis of the app is AI learning and urination behavior recognition, and various ML and deep learning techniques were used for this purpose. Figs. 1 and 2 show the research process and concepts.

Researchers at Sejong Chungnam National University Hospital are working on a technology that utilizes image data from microendoscopy, computed tomography (CT), etc. to assist in the diagnosis of conditions such as ureteral stricture and urinary stones [12]. Research on diagnostic aid for ureteral strictures is important because it is often difficult to identify the area of stricture during surgery, making it difficult to remove the stricture. For decision support, AI technology is used to automatically recognize the area of stricture in ureteral images obtained through microendoscopy to assist clinicians during surgery. The AI technology used utilizes deep learning techniques such as ResNet-50 to provide recognition information in real time [13-17]. Fig. 3 shows a concept map of this study. In addition, since accurate diagnosis is paramount for the surgery and treatment of urolithiasis, research is also underway to support the diagnosis of urolithiasis by utilizing abdominal CT and Carm imaging data. To support the diagnosis of urinary stones, research is underway on a technology that automatically diagnoses urinary stones and evaluates their condition based on image data [12]. For this purpose, research is being conducted to explore the optimal technology method in terms of image processing (Watershed), machine learning (support vector machine), and deep learning (ResNet-50, Fast-RCNN [18]). Fig. 4 shows a concept map of this research.

In addition, researchers at Seoul National University Bundang Hospital developed a mobile urethral flow measurement technology through acoustic signal analysis. In 2017 and 2018, they conducted an internal performance evaluation of the urinary flow measurement technology through clinical trials to confirm its validity software (SW) with this technology is an acoustic analysis-based urinary flow measurement software, which was registered as a class 2 medical device with the U.S. Food and Drug Administration (FDA) in 2020. The technology measures urination patterns based on acoustic analysis and AI, enabling noninvasive testing with a smartphone. Based on the sound produced when urine hits water, it automatically records urination volume, peak flow rate, urination time, etc. and displays the data in a data graph and stores the measurement data in a dashboard. AI technologies such as linear/nonlinear regression analysis, classification analysis, and deep learning have been applied here. Fig. 5 shows the workflow of this device.

Neurogenic bladder refers to overactive bladder, a condition in which patients diagnosed with abnormalities in bladder storage function by dynamic testing, such as stroke patients, develop bladder overactivity due to nerve damage. Existing studies [19, 20] have reported that stroke affects the symptoms or disease presentation of neurogenic bladder. There are several of these patterns, among which the pattern of facial paralysis is easy to identify. The asymmetry criterion of the face is mainly important, and precise detection of it is essential to prevent stroke. In this study, we analyze the early symptoms of stroke through ML-based recognition [21] of facial images collected by detecting facial paresis symptoms in stroke patients.

In addition to providing various light sources, they also provide various image format conversion technologies required for diagnosis. In addition, as it targets the elderly, it is possible to install AI analysis software that can accurately diagnose the early stages of stroke so that it can be utilized in daily life. The experimental test environment proposed in this paper is as follows: a software that analyzes images obtained from face and arm motion videos and displays suspected stroke images based on AI learning (object). Among the analyzed results, it supports the function of detecting and capturing objects in the acquired face and arm motion images, which can be used as reference indicators for stroke diagnosis. Among the video frames, still cuts are taken by selecting frontal shots, facial landmarks, eyebrows and eyes, corners of the mouth, and lip frames. During the experiment, one of the precursor symptoms of stroke patients is that the face is not symmetrical from side to side due to paralysis on one side of the face. In order to use these features to distinguish patients from normal people in the input image, we need to extract features from 2 groups of face databases to learn the difference. We need to detect the angle and tilt to calculate the distance between feature points in the face region and compare the left and right differences to calculate the degree of asymmetry. Since existing databases have different sizes and tilts of face regions, we need to go through a preprocessing process to normalize the size and angle of the detected faces. Then, to obtain the rotation angle, we store the left effect values of the left and right end landmarks as the rotation angle. We store an affine transformation matrix to align the image using this rotation angle. After applying the angle to extract landmarks from the rotated image, calculate the distance values to find features for left-right symmetry analysis. To extract the calculated distance values into meaningful data, we divide and compare the left and right distances and store the larger value, and also extract calculated values such as slope, angle, and ratio of eyebrows and lips. The extracted data is then normalized, labeled, and trained with a convolutional neural network (CNN) model.

They found that the training accuracy averaged 86% and the validation accuracy averaged 83%, and the accuracy of the control and patient groups was also close to 80%. We recorded the accuracy and loss as a function of the number of epochs, and found overall similar performance on the training data. The results are shown below (Tables 1, 2). Specifically, the classification accuracy of the patient data was 88% and the control data was 87%, which can be seen in the following tables (Tables 3, 4). In the end, we found that using AI to diagnose stroke quickly and accurately had an indirect, if not a direct, positive effect on the neurogenic bladder.

This is the second study [22] of neurogenic bladder, and it is a case of judgment based on speech data rather than images. In the case of speech, this includes inaccurate pronunciation or unnatural intonation, and accurately detecting these patterns is crucial for stroke prevention. In this research case, we propose a platform that can precisely analyze the speech of stroke patients with neurogenic bladder and detect them in advance to prevent the disease. Digital therapy in the field of neurogenic bladderrelated stroke management is a type of software therapy that changes a patient’s behavior or lifestyle through a series of digital stimuli such as images and sounds, and collects and analyzes data accordingly. Currently, the application of digital therapies is gradually expanding to fields related to behavior and habit change, such as mental illness and chronic diseases. Digital therapies can serve as a complement to conventional medications in conditions that require ongoing management. The SW proposed in this study can provide scientific monitoring feedback to users or surrounding nursing personnel by iteratively learning and analyzing the utterance data of stroke patients and expressing continuous feedback and interest. The SW in this study is an AI SW that acquires specific sentence speech data of stroke clinical patients and analyzes it to warn of early stages of neurogenic bladder-related stroke. It extracts feature values by analyzing the input speech and converts them into stroke indexes and stores them. The stored stroke index is classified into learned stroke indicators and finally displays whether a stroke has occurred. The input speech is processed for feature analysis with the mel-frequency cepstral coefficient (MFCC) algorithm, and a CNN learning algorithm is used to extract the speech disorder index and derive classification values by index. The MFCC algorithm is used to extract feature data from speech to learn the differences between patients and normal people.

Fig. 6 shows the main feature spectrum of the MFCC algorithm, which is the result of the feature vectorization process based on the input voice data, and finally, through voice stroke symptom detection, risk factors can be identified in advance and notification services can be provided, and the user’s results can be stored in the DB and used for retraining.

Fig. 7 shows the data processing flow chart for performance evaluation. It tested the analytical model by processing 30 abnormal data and 30 normal data in real time. The results showed high test accuracy of 98.7% for normal data and 99.6% for abnormal data. Table 5 shows the final test results.