Automated measurement of penile curvature using deep learning-based novel quantification method

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Urology

Background:

• Penile curvature (PC) assessment is crucial for diagnosing conditions like Peyronie's disease.
• Current measurement methods can be subjective and lack precision.
• Automated, objective tools are needed for reliable PC quantification.

Purpose of the Study:

• To develop and validate a deep learning-based automated method for accurate penile curvature measurement using 2D images.
• To improve the precision and reliability of PC assessment in clinical practice and research.

Main Methods:

• A dataset of 913 images of 3D-printed penile models with varying curvature was generated.
• YOLOv5 and UNet models were used for penile region localization and segmentation.
• An HRNet model was trained to predict landmarks and calculate curvature angles.
• The method was validated on real patient images against expert assessment.

Main Results:

• The automated method achieved a mean absolute error (MAE) of <5° on penile models.
• For real patient images, AI predictions showed a deviation of 1.7° to 6° compared to clinical experts.
• Accuracy was maintained across different degrees of penile curvature.

Conclusions:

• The developed deep learning approach offers a reliable and automated solution for measuring penile curvature.
• This method has the potential to significantly enhance patient assessment for surgeons and hypospadiology researchers.
• It overcomes limitations associated with conventional PC measurement techniques.