Attention: Restrictions on use of AUA, AUAER, and UCF content in third party applications, including artificial intelligence technologies, such as large language models and generative AI.
You are prohibited from using or uploading content you accessed through this website into external applications, bots, software, or websites, including those using artificial intelligence technologies and infrastructure, including deep learning, machine learning and large language models and generative AI.

How Pixel Segmentation Can Be the Future of Hypospadias Phenotyping

By: Nicolas Fernandez, MD, PhD, FACS; Adam Maxwell, PhD; Lauren Erdman, MSc | Posted on: 04 Jan 2023

Congenital anomalies are described as a variation from a functional and normal phenotype. Hypospadias is no exception. The evolution of hypospadiology has been centered around restoring the anatomy into a more natural and functional phenotype. As such, all efforts made have been on developing and improving technical surgical principles and concepts to master the reconstruction of individuals born with this birth defect. Despite multiple technical advances, there is no single procedure which results in a reproducible and predictable outcome. Prolific surgeons develop an unconscious artistic ability to interpret the phenotype and perform successful reconstructions. This subjectivity has been recognized as a major limitation in hypospadias phenotyping, and for decades surgeons have attempted to overcome this issue by incorporating anthropometric measurements into their assessment of the phenotype. In general, variables such as glans and urethral plate width, ventral curvature, and meatal location have been accepted as predictive variables. Nonetheless, these attempts at increasing objectivity have not resulted in consistently good prediction of surgical outcomes, with the current classifications still having a significant subjective component, particularly when the urethral plate is described. Maybe it is time to stop and redefine how we see the hypospadias phenotype and evaluate how the urethral plate can be described using novel technologies.

Figure. Heat map output for pixel cluster k-mean analysis.

A major step in the evolution of surgical reconstruction has been the concept of preserving the urethral plate and incorporating it as part of the reconstruction. Widely accepted surgical techniques such as the meatal advancement glansplasty, Matheiu, tubularized incised plate, and Thiersch Duplay include the urethral plate as the main structure for reconstruction. As an artistic process, surgeons use their skills and knowledge subjectively to evaluate the urethral plate and make a decision about the most suitable technique for the patient. They also define the boundaries and how to use the plate for the reconstruction. Nonetheless, there is a lack of understanding of how an experienced surgeon makes that decision and how that translates to selected tissue quality and wound healing. Histological architecture of the urethral plate and surrounding tissues has been previously explored but it has never been correlated to a standardized phenotyping with subsequent postoperative follow-up.

Reports of urethral plates with abnormal or pathological histology have been described and identified by ongoing research in our group. Patients with abnormal pathological urethral plate architecture might not have adequate healing, which can impact wound healing independent of the surgical technique. A preliminary ongoing study using histology tissue mapping has demonstrated a 20% proportion of abnormal histology findings confined within the urethral plate. The most relevant finding has been chronic inflammatory lymphocyte infiltration in the subepithelial region. Recognizing the limitation of not being able to have a priori preoperative information before histology analysis is made, we have decided to explore the use of novel pixel analysis technologies to characterize the urethral plate, trying to reduce subjectivity and explore its predictive potential to identify abnormal histological findings.

Based on prior experience using computer vision methods to reduce subjective hypospadias phenotyping, we are currently using pixel cluster analysis to describe and predict the correlation to histological findings (see Figure).1,2 The current pilot analysis has proven that pixel clustering can discriminate between different urethral plates when compared to the Glans, Meatus, Shaft score. Using principles demonstrated in dermatology and dermoscopy research for digital image analysis, we plan to extrapolate these concepts to our methodology.3,4 Superpixel segmentation using linear iterative clustering can help identify pathological skin lesions subsequently confirmed on histological analysis. Other image features such as entropy or texture analysis are approaches that may be informative and will help with urethral plate phenotype description and classification.4

Even the more traditional computer vision techniques described above represent a dramatic step forward for objective hypospadias phenotyping. Therefore, we are hopeful that newer, more flexible machine learning methods for computer vision, such as convolutional neural networks (CNNs), will make reliable hypospadias phenotyping easier. These methods haven’t been used up to this point due to the requirement for very large samples of images (10K-100K); however, transfer learning has more recently been shown to produce high performing CNNs with far fewer samples needed. CNNs have also been criticized for their lack of interpretability; however, vision transformers have been proposed as a special case of CNNs which learn attention maps and focus on specific areas of the image. These attention maps are interpretable as they produce a heat map over the image, enabling a human to identify which part of the image is most important for the downstream prediction. Importantly, transformers are not data-type specific and indeed were first developed for speech. Therefore, transformer models can also be used to encode other data types such as clinical notes, sequencing data, and more.

Although we see a promising future for machine learning in the field of hypospadias, we cautiously recognize some limitations.2 The prevalence of the condition makes it difficult to build a database that is large enough to create and test a neural network algorithm. The complexity and variability of the surgical reconstruction generates a challenging confounding factor that needs to be considered when analyzing data. To mitigate and overcome this barrier, we plan to collect additional variables that can be informative for the algorithm and create a more comprehensive patient-centered phenotype, which will provide a more stable target or input to any models built. This will include epidemiological, anthropometric, histological, and genetic variables. This will only be possible if standardized clinical and surgical approaches are implemented for hypospadias repair. Efforts at Seattle Children’s Hospital have allowed our group to develop a standardized practice amongst providers. Current ongoing data collection and follow-up are underway.

  1. Fernandez N, Lorenzo AJ, Rickard M, et al. Digital pattern recognition for the identification and classification of hypospadias using artificial intelligence vs experienced pediatric urologist. Urology. 2021;147:264-269.
  2. Khondker A, Kwong JCC, Malik S, et al. The state of artificial intelligence in pediatric urology. Front Urol. 2022;
  3. Winkler JK, Fink C, Toberer F, et al. Association between surgical skin markings in dermoscopic images and diagnostic performance of a deep learning convolutional neural network for melanoma recognition. JAMA Dermatol. 2019;155(10):1135-1141.
  4. Annaby MH, Elwer AM, Rushdi MA, Rasmy MEM. Melanoma detection using spatial and spectral analysis on superpixel graphs. J Digit Imaging. 2021;34(1):162-181.