TL;DR
This paper introduces VCUG-DAM, a novel deep learning system that automates the detection and diagnosis of urinary tract abnormalities from Voiding Cystourethrography (VCUG) images. By dynamically weighting the importance of multiple images per patient, VCUG-DAM achieves high accuracy and significantly enhances the diagnostic capabilities of clinicians.
What Problem Does This Paper Address?
Voiding Cystourethrography (VCUG) is a crucial fluoroscopic technique for diagnosing various urinary tract conditions in children, particularly vesicoureteral reflux (VUR). However, interpreting VCUG images is challenging due to variations in technique and interobserver variability, leading to inconsistent diagnoses. While deep learning (DL) has shown promise in medical imaging, existing DL models for VUR often focus solely on grading and require manual image selection, limiting their clinical utility. This paper addresses the need for a comprehensive, automated system that can not only diagnose VUR but also segment and diagnose other urinary tract abnormalities (bladder, urethra) from multiple VCUG images, while also improving clinician performance.
Methods
The core of this work is the VCUG-DAM (Voiding Cystourethrography Diagnostic Artificial Intelligence Model). It employs a novel architecture with two main components:
- Image-level Network: This component uses a unified transformer-based architecture (initialized with ViT-Base weights pre-trained on ImageNet) to simultaneously perform classification and segmentation. It takes a single VCUG image and outputs lesion masks for the bladder, urethra, and ureter, along with predicted grades for bladder abnormalities, urethral abnormalities, and Left/Right VUR.
- Patient-level Network: This model takes the features extracted from the image-level network for all relevant VCUG images of a patient. It uses a dynamic multi-image weighting mechanism (specifically, a multitask attention layer) to assess the contribution of each image to the overall diagnosis. This allows the model to dynamically determine the importance of each image for a comprehensive patient-level diagnosis.
The system is designed to overcome limitations of previous DL approaches by:
- Handling multiple VCUG images per patient.
- Dynamically assessing image importance.
- Performing multitask diagnoses (segmentation and classification of bladder, urethra, and VUR).
Datasets
The study utilized a large, multicenter dataset:
- Training and Internal Validation: Data was collected from the Children's Hospital of Fudan University, comprising 1,270 VCUG examinations with 5,644 images. This was split into a training set (889 VCUGs) and an internal validation set (381 VCUGs) using a 7:3 ratio.
- External Validation Dataset 1: 256 VCUG examinations (1,694 images) from 10 hospitals were used for external validation and to assess the impact of VCUG-DAM on clinician performance.
- External Validation Dataset 2: 134 VCUG examinations (490 images) from 4 hospitals were used to evaluate the consistency of VUR grading with VCUG-DAM assistance.
In total, 7,899 VCUG images from 1,660 patients across 15 Chinese hospitals were collected between 2021 and 2023.
Key Insights and Ideas
- Dynamic Multi-Image Weighting is Crucial: The ability to dynamically assess and weight the importance of different VCUG images from the same patient is key to achieving accurate patient-level diagnoses. This moves beyond single-image analysis and captures the temporal/morphological changes across an examination.
- AI as a Clinician Augmentation Tool: VCUG-DAM doesn't just aim to replace clinicians but to significantly enhance their diagnostic performance and consistency, particularly for junior clinicians.
- Comprehensive Diagnosis Beyond VUR: The model's ability to simultaneously segment and diagnose bladder and urethral abnormalities alongside VUR grading offers a more holistic diagnostic capability than previous VUR-specific DL models.
- Interpretability through Attention Maps: The use of attention maps provides insights into which image regions the model focuses on, aiding in understanding its predictions and potentially highlighting critical diagnostic areas for clinicians.
Key Results
- High Model Performance:
- Image-level: Achieved accuracies of 0.994 for bladder, 0.9723 for urethra, and AUCs of 0.9568 (left VUR) and 0.9249 (right VUR).
- Patient-level: Achieved AUCs of 0.8772 for bladder, 0.7752 for urethra, 0.9443 for left VUR, and 0.9342 for right VUR on the internal validation dataset.
- Clinician Performance Improvement: VCUG-DAM assistance significantly improved clinician diagnostic performance:
- Mean AUCs increased from 0.8185 to 0.9456 for bladder.
- Mean AUCs increased from 0.6507 to 0.7943 for urethra.
- Mean AUCs increased from 0.6288 to 0.9641 for left VUR.
- Mean AUCs increased from 0.7305 to 0.9506 for right VUR.
- All improvements were statistically significant (P < 0.0001).
- Improved Inter-Clinician Agreement: VCUG-DAM assistance also improved the consistency of VUR grading among clinicians, as measured by kappa values.
Limitations and Future Work
- Ground Truth Variability: While efforts were made to ensure consensus, ground truth labels derived from expert readings can still have inherent variability.
- Data Imbalance: The dataset had an insufficient number of normal bladder and urethral samples. The authors addressed this using a multitasking learning strategy with a Vision Transformer architecture.
- Recall Bias: Despite a washout period and randomized ordering in reader studies, recall bias could not be entirely eliminated.
- Lack of Detailed Discordance Analysis: The study did not perform detailed comparisons of discordance between VCUG-DAM and clinicians.
- Prospective Studies Needed: The authors emphasize the need for prospective studies with larger sample sizes to further validate the tool's clinical applicability and refine it. They look forward to testing and refining VCUG-DAM in such settings.
Summary updated