INTRODUCTION
Urodynamic studies are diagnostic procedures that evaluate the function of the lower urinary tract by measuring pressures and flows within the urethra and bladder. The Good Urodynamic Practices (GUP) aim to standardize these procedures and interpretations to ensure accurate and reliable results, thereby providing a scientific basis for clinical diagnosis and treatment, which was used by the International Association of Urinary Control (International Continence Society, ICS) to develop a series of measurements of urodynamics, quality control mechanisms, and technical specifications of the results [1, 2].
The changes of urodynamic signal features are complex, and traditional machine learning cannot identify the signal features in real time, while deep learning (DL) has the advantages of capturing high-dimensional image features and strong data processing capabilities, which can just complete the task [3]. DL algorithms are a subset of machine learning that uses multilayer neural networks to simulate how the human brain processes information to learn complex patterns and features from large amounts of data [4]. It has unique advantages in medical image processing, such as automatic feature extraction, batch data processing, generalization, and multitask learning. The implementation of the previous urodynamic technical specifications required the operator to complete the urodynamic examination and pay attention to the changes of the urodynamic signals in real time, which was often difficult for inexperienced operators, and eventually the technical specifications were not strictly implemented or ignored, resulting in the unreliable results of the urodynamic report [5]. Though the technical specification of urodynamics is clear and objective, it is often affected by subjective factors such as the difference of operator level in the implementation process, which leads to its popularization and application being restricted. Therefore, our study developed a real-time urodynamic signal recognition system to accurately identify signals with objective and unified standards to reduce the influence of the above subjective factors, thereby improving the quality of urodynamic examination in different regions and reducing healthcare disparities in underserved areas.
MATERIALS AND METHODS
Patient Cohort
Urodynamic image data from 2 cohorts of patients with neurogenic bladder were studied. The first cohort was made up of 300 adult patients with neurogenic bladder in our center from 2012 to 2018 and used to train and validate the DL model to determine performance metrics. Data included 1,960 images with urodynamics technical specifications related to typical signal, cough wave (grades A, B, and C), detrusor overactivity (DO), rectal peristalsis, and abdominal muscle contractions. Poor-quality images were excluded. The second cohort was made up of 100 patients with urinary dysfunction in our center from 2020 to 2021 used to test the depth of the DL model. The baseline characteristics of patients are shown in Table 1. A total of 695 images were included. The flow chart of the study process is shown in Fig. 1.
Typical Urodynamic Signal Pattern Category
The image study mainly included 4 channels (urodynamics abdominal pressure [Pabd], vesical pressure [Pves], detrusor pressure [Pdet], and flow of urodynamics), which included 6 kinds of typical signals and cough waves (grades A, B, and C), DO, rectum peristaltic waves, and abdominal muscle contractions. The most common method of cough wave evaluation with stress signaling quality was used [6, 7] (cough grade A: smaller wave/bigger wave 70% or higher; cough grade B: >70% smaller wave/bigger wave >30%; and cough grade C: 30% or greater smaller wave/bigger wave). Grade A was acceptable, while grades B and C were not. DO waves showed synchronous changes in Pves and Pdet but no or insignificant changes in Pabd. Rectal peristalsis waves showed the positive wave in Pabd, the negative wave in Pdet, and no significant change in Pves. Changes in abdominal muscle contraction wave Pabd and Pves, Pdet were mild. All signal examples are shown in Fig. 2B.
Data Preprocessing
All urodynamic studies were completed on a Laborie machine. The image study used local screenshots of urodynamics tests with a size of 512×512 pixels after considering the urodynamics used the process of training data image characteristics and the actual application. The image annotation was completed by 2 doctors with more than 5 years of working experience, and was included in the database (including development cohort and test cohort) after being corrected by one doctor with more than 10 years of working experience. The data-enhancement approach included the translation of all types of typical signals only, and no patient personal information was included. Labeling software was used for data annotation. The label data format was YOLO (you only look once), the label file format was txt, and the image format was png. The artificial intelligence system detection process is shown in Fig. 2.
DL Algorithms
This study applied yolov5 (https://github.com/ultralytics/yolov5) 5 series model of the (n, s, m, l, x) training test, at least within n model parameters, for the fastest but lowest-precision model parameters. The model of depth and width increased the ability to perform increased-complexity image feature extraction and gradually improved accuracy, but slowed the reasoning speed. The real-time recognition model of typical urodynamic signal by the yolov5 algorithm was combined with real-time screen capture (size 512×512 pixels) to assess. The detection mode is shown in the supplemental video. After classifying each type of signal, gradient class activation map (GradCAM) class activation graph-visualization analysis [8] was used to aid understanding of the focus area for the model classification of consistency with physicians.
Statistical Analysis
For training and validation, in accordance with the training set (70%) and validation set (30%), we determined the proportion of model training and performance evaluation in the validation set. Performance indicators included the recall rate (sensitivity), precision rate, F1 score, precision-recall (PR) area under curve (average precision, AP), and mean AP (mAP). All DL experiments were done using PyTorch in Python 3.8.
RESULTS
Typical Signal Pattern Recognition Analysis
PR curve is the curve enclosed by the precision rate values of the model under different recall rates, and the AP value was the AP of each signal type, which is the area under the curve value of PR curves. The quality of cough response signal was classified into 3 grades according to the most commonly used standard. Grade A represented high‐quality pressure transmission, and grades B and C represented low‐quality pressure transmission. The best AP value of cough grade A was 0.89, and the best grades B and C value was 0.71 and 0.86. The best AP value was 0.89 for DO signals. The best mAP rectal peristalsis signal value was 0.86, and the best mAP value of abdominal muscle contraction signal was 0.83.
Typical Signal Prediction Through Urodynamic Traces
The results of the yolov5 model after 250 epoch of training in the training set and validation of the validation set are shown in Table 2. And, the results of the test set are shown in Table 3. As Table 3 shows, the detection performance of yolov5l yielded the highest comprehensive index scores (F1, 0.84; mAP, 0.89). The F1 score, a comprehensive precision evaluation, was negatively correlated with the false-positive rate, the recall rate was negatively correlated with the nonresponse rate index, and the mAP was the average of the model prediction of accuracy for all categories. The verification of the yolov5 model on the test set showed similar performance. Yolov5l had the best detection performance and the highest comprehensive index score (F1, 0.81; mAP, 0.83) and inference speed was 10.2 msec per image (Table 3).
The classification accuracy of the yolov5 model for each signal is shown in Table 3. The recognition accuracy of yolov5l for the cough A, cough C, and DO signals was above 95%, and rectal peristalsis was close to 95%. The accuracy of cough B and abdominal muscle contraction was low (mAP, 0.67, 0.75, respectively) due to the partial overlap among signal features.
Examples of multiclass typical urodynamic signals is shown in Fig. 3A. The GradCAM class activation heatmap was used to visualize the focus areas of different signals identified by the yolov5 model. The focus areas of different signals of the model (the darker the color of the area, the greater the weight) were basically consistent with those of urodynamic experts (Fig. 3B). The training process of yolov5 is in Fig. 4.
DISCUSSION
Urodynamic quality control is categorized as qualitative or quantitative analysis. Qualitative analysis mainly consists of typical signal pattern recognition (e.g., DO, cough test signal, rectal peristalsis), such as the establishment of the quantitative analysis in a typical value range, clear-urine baseline value range of the urodynamic trace [1, 2], The ICS guidelines include different criteria to use for different examination positions [1, 2], but the criteria can be changed according to differences at urodynamic centers. At the same time, the guidelines point out that artifacts in urodynamic reports are often found in retrospective analysis but cannot be corrected, which reflects the importance and necessity of real-time quality control but has not been shown to be of high quality.
In our study, the model monitors the quality of urodynamic signal transmission through real-time grading evaluation of cough test signals, and then improves the quality of urodynamic examination. Recognition of DO waves can assist in the diagnosis of overactivity bladder, and interference signals such as rectal peristaltic wave and abdominal muscle contraction wave can be used to improve the recognition accuracy of DO waves, improve interpretation of urodynamic, and more directly and conveniently observe the vesicourethral function of patients and evaluate the relevant urodynamic changes. Assist clinicians to efficiently complete urodynamic report and realize intelligent urodynamic examination. The real-time urodynamic signal recognition system can play the role of clinical assistant, reduce the difficulty and burden of testing, and promote and implement the urodynamic technical specifications. Currently, machine learning algorithms are widely used in the field of urodynamics [9]. Some research teams that have carried out relevant research on the combination of machine learning and urodynamic examination have achieved high detection accuracy (i.e., the detection of DO signals and bladder compliance assessment) [10-12]. The DL algorithm is used to analyze the urine stream of sound changes, assess the patient’s urine flow rate, increase convenience, and confirm that DL can improve the exploration of urodynamics [13]. However, most of the model is mainly used for urodynamic test-image post-processing or single signal recognition and depends on the quality of the urodynamic test. High-quality examination, however, require the process to strictly follow the GUP, which is difficult for some physicians with little experience. Therefore, this study proposes a scheme combining a DL algorithm for urodynamic real-time signal recognition, which helps operators complete urodynamic examinations in accordance with GUP and reduces the difficulty of the work.
In addition to the role of quality control, the model’s recognition of typical signals is helpful for the auxiliary diagnosis and analysis of lower urinary tract dysfunction. First, the model can identify signals throughout the entire examination process, and the diagnostic accuracy of DO waves is over 95%, which aids in identifying symptomatic DO waves during examinations and in diagnosing overactive bladder. A multicenter study also confirmed the importance of quality control in the diagnosis of bladder hyperactivity [14]. At the same time, the model can identify the signal of abdominal muscle contraction, which can help the operator more rapidly capture the signal of Pabd urination and change in urine flow rate, which helps in the diagnosis of lower urinary tract obstruction [15]. This study tested yolov5 in all types of models and compared their performance to select the most suitable for clinical deployment of the model of signal and found the optimum properties of yolov5l. Yolov5x had comparable performance but more model parameters requiring more computing resources. Therefore, the model parameters and structure depth of yolov5l are more compatible with the task requirements of real-time quality control, which provides a reference for later clinical deployment and application.
Deploying the system into the real clinical workflow will exert its greater clinical value, but there are some potential challenges. First of all, there are differences in the equipment of urodynamics. Our system is developed based on the urodynamic equipment of Laborie. Whether other equipment can achieve the same effect needs further verification to exclude the influence of equipment differences. Secondly, clinical operators required unified and standardized software operation training and urodynamic practice training, so as to improve the use efficiency and inspection quality. Finally, our team has made the system into a computer application, with a simple operation interface to help overcome the challenges, which can be deployed in urodynamic equipment for clinical application. There are also some limitations in this study. The model test was completed in a single center, and no multicenter test has been conducted to verify the generalization ability of the model. We will verify and improve it in the next stage of our research.
In conclusion, our research has developed a real-time urodynamic signal recognition system to assist clinicians in quality control during examinations. This system aims to enhance the standardization and quality of urodynamic assessments, reduce regional disparities in medical services, and benefit more patients with lower urinary tract dysfunction.
PDF Links
PubReader
ePub Link
XML Download
Full text via DOI
Download Citation
Print
