Dynamic Multi-Image Weighting for Automated Detection and Diagnosis of Abnormal Urinary Tract on Voiding Cystourethrography with a Deep Learning System: A Retrospective, Large-Scale, Multicenter Study

Abstract

We aimed to develop a voiding cystourethrography (VCUG) diagnostic artificial intelligence model (VCUG-DAM), which relies on a novel architecture to automatically segment and diagnose the bladder, urethra, and ureters using a single VCUG image, while dynamically assessing the relative importance of each image. A total of 7,899 VCUG images from 1,660 patients across 15 Chinese hospitals were collected between 2021 and 2023. In stage 1, we assessed the performances of the VCUG-DAM model. The patient-level area under the curve (AUC) of VCUG-DAM was 0.8772, 0.7752, 0.9443, and 0.9342 for bladder, urethral, left, and right vesicoureteral reflux (VUR), respectively. In stage 2, we explored whether the VCUG-DAM model could improve the diagnostic ability of clinicians. VCUG-DAM improved the clinician's diagnostic performance, with mean AUCs increasing from 0.8185 to 0.9456 for the bladder, 0.6507 to 0.7943 for the urethra, 0.6288 to 0.9641 for the left VUR, and 0.7305 to 0.9506 for the right VUR (all P < 0.0001). In stage 3, the consistency of the VCUG-DAM for VUR grading was validated. VCUG-DAM improved inter-clinician agreement for VUR grading. The fully automated VCUG-DAM demonstrated high accuracy, reliability, and robustness in multitask diagnoses of urinary tract abnormalities across multiple VCUG images, while improving the diagnostic ability of clinicians as an auxiliary tool.

Fig. 1.

Confusion matrices and ROC curvature of the VCUG-DAM performance at the image level (A to D) and patient level (E to H) in the internal dataset. L-VUR, the grading of the left VUR; R-VUR, the grading of the right VUR.

Fig. 2.

Performance of the VCUG-DAM model and clinicians with and without model assistance in classification. (A) Urethra. (B) Bladder. (C) Left-VUR. (D) Right-VUR. AI, VCUG-DAM; C1, mean junior clinicians; C2, mean attending clinicians; C3, mean senior clinicians; R1, mean junior radiologist; R2, mean attending radiologist; R3, mean senior radiologist. There were 2 senior clinicians, 2 attending clinicians, and 2 junior clinicians or radiologists in each group. The blue dots denote the performance of the clinicians without assistance from the VCUG-DAM model. The gray dots denote the performance of the clinicians with assistance from the VCUG-DAM model.

Fig. 3.

The performance of VUR grading the clinicians with and without the VCUG-DAM assist on the external validation dataset 2. Left-VUR without model assistance (A) and with model assistance (B). Right-VUR without model assistance (C) and with model assistance (D).

Fig. 4.

Three VCUG images selected from the testing set (left) with manual segmentation (second column), automated classification and segmentation (third column), and attention map of the prediction model (right). The 3 colors in the predicted label represent 4 classes of image-level prediction including bladder (red), ureter (yellow), and urethra (green). AI, artificial intelligence.

Fig. 5.

Image weight map of a patient with VUCG examination shown as an example. The first row (raw photos) shows the original VCUG image of a patient. The second row shows the segmentation and diagnosis of a single image of the VCUG-DAM model. The third row (ground truth) shows the segmentation of the VUR, urethra, and bladder on each VCUG image. The fourth row shows the results of merging the segmentation on the original VCUG images. The fifth row shows the calculated contributions to the VUR, urethra, and bladder for each VCUG image. The final row shows the prediction of VCUG-DAM on the patient level.

Fig. 6.

Flowchart of the study population.

Fig. 7.

Overview of the study design. (A) Data collection. (B) Development of VCUG-DAM models for the image level and patient level on the VCUG image. (C) Twelve clinicians without and with the assistance of VCUG-DAM to diagnosis. (D) Model analysis. Center A is Children’s Hospital of Fudan University; Center B consists of 10 hospitals; Center C consists of 4 hospitals.