SPU Main Site  |  Past & Future Meetings
Society For Pediatric Urology

Back to 2022 Abstracts


Comparison of hypospadias phenotyping using machine learning pixel segmentation to GMS score.
Nicolas Fernandez, MD, PhD, FACS1, Joshua Caldwell, MD2, Meghana Noonavath, MD2, Adam Maxwell, PhD2.
1Seattle Children's Hospital, Seattle, WA, USA, 2University of Washington, Seattle, WA, USA.

BACKGROUND: Assessment of hypospadias phenotype determines whether the anatomy is favorable for reconstruction. Despite its subjectivity, the GMS score has been used as a standardized tool to define the severity of the phenotype to predict outcomes. The use of digital image analysis has proven to be feasible and prior efforts by our team have demonstrated that machine learning algorithms can emulate an expert's assessment of the phenotype. Nonetheless, creation of these image recognition algorithms involves a subjective input. To reduce subjectivity in phenotype evaluation, we propose a novel approach to analyze anatomy using digital image pixel analysis and subsequently compare results using the GMS score. Our hypothesis is that pixel cluster segmentation can discriminate between favorable and unfavorable anatomy.
METHODS: A total of 148 digital pictures of patients with different types of hypospadias were classified by 3 independent experts following the GMS score into "favorable" (54), "moderate" (44) and "unfavorable" (50) glans. For each image, the region of interest was segmented separately by two trained evaluators into "glans," "urethral plate," "foreskin" and "periurethral plate." The values obtained for each segmented region using machine-learning statistical pixel k-means cluster analysis were analyzed and compared to the GMS score given to that image using an ANOVA analysis.
RESULTS: Analysis of 592 segmented images demonstrated that k-means pixel cluster analysis discriminated "favorable" vs "unfavorable" urethral plates. One-way ANOVA comparing k-means data between "favorable," "moderate," and "unfavorable" groups revealed a statistically significant difference in at least one cluster (out of four) for each segmented region: Glans (k-mean for favorable (7.28), moderate (9.45), and unfavorable (6.77), p=0.0046, p=0.045), urethral plate (k-mean for favorable (9.13), moderate (11.38), and unfavorable (8.61), p=0.0062, p=0.020), foreskin (k-mean for favorable (8.04), moderate (11.16), and unfavorable (8.75), p=0.0073), and periurethral plate (k-mean for favorable (7.087, 7.435), moderate (10.60, 9.81), and unfavorable (7.70, 7.62), p=0.00040, p=0.0018). Further, statistical significance was demonstrated in one-way ANOVAs comparing the absolute difference in k-means data between the urethral plate and foreskin (favorable (4.35), moderate (4.08), unfavorable (2.96), p=0.013), between the urethral plate and glans (favorable (4.89), moderate (4.30), unfavorable (6.41), p= 0.021), and between the urethral plate and normal skin (favorable (6.53), moderate (8.62), unfavorable (4.80), p=0.0065).
CONCLUSIONS: Digital image segmentation and statistical k-means cluster analysis can discriminate anatomical features, performing similarly to GMS score in characterizing hypospadias phenotype. Future applications of this approach will serve to build more complex deep learning algorithms.


Back to 2022 Abstracts