AI-driven 3D assessment for Peyronie’s disease

Peyronie’s disease (PD) is a fibrotic disorder of the tunica albuginea resulting in penile curvature and other deformities during erection, often causing significant functional and psychological distress. Despite its prevalence, current clinical assessment methods remain highly variable and prone to inaccuracy.1 In-office evaluation typically involves pharmacologically induced erections and manual goniometer measurements, but the absence of standardized measurement protocols leads to considerable inter- and intra-observer variability. Digital photography offers a less invasive alternative and enables at-home assessment during natural erections; however, these methods rely on two-dimensional images, which cannot capture the complex three-dimensional morphology of penile deformities, particularly in multiplanar deformities such as indentation or hourglass. Three-dimensional representation of the penile shaft could overcome these limitations by providing accurate morphological information, enabling remote expert consultation, surgical outcome simulation, and improved training for urologists. Existing 3D capture methods require expensive scanning equipment or extensive image sets from multiple angles, restricting their practical use. Advances in artificial intelligence and 3D computer vision make it possible to reconstruct 3D objects from a small number of photographs, creating an opportunity for PD assessment on 3D representation.2 We present a fully automated pipeline that reconstructs a 3D model of the penile shaft from as few as three 2D photographs, using a generative model for reconstruction and an AI-based assessment algorithm for curvature measurement. This approach offers a reproducible, standardized, and objective method for PD evaluation that eliminates the need for manual goniometry and overcomes the depth limitations of 2D imaging. By bridging AI-driven 3D reconstruction with automated PD assessment, this method has the potential to transform both clinical and at-home assessment of PD. To address the absence of publicly available datasets for PD with controlled imaging conditions for 3D reconstruction, we developed a fully synthetic dataset. Thirty 3D-printed penile models, including deformities such as double curvature, hourglass, and indentation, as well as controls, were photographed from standardized multi-view angles using a smartphone. In parallel, 182 digital computer-aided design (CAD) models were automatically generated to systematically represent varying angulation, curvature location, direction, and circumcision status, each rendered from multiple standardized views. Ground truth 3D representations were retained for performance benchmarking (see samples in Supplementary Figure S1). Our automated inference pipeline integrates four stages: (1) image preprocessing to isolate and align the penile shaft, (2) 3D mesh reconstruction from multi-view images using a state-of-the-art generative model (TRELLIS3), (3) standardized rendering of the reconstructed model, and (4) curvature analysis using an adapted version of the AI-based PenoMeter4 tool. This process enables fully automated PD assessment, generating a 3D representation without specialized scanning hardware (Supplementary Figure S2). Performance was benchmarked using controlled experiments on both printed and digital models using an evaluation pipeline (Supplementary Figure S3), quantifying reconstruction quality and evaluating curvature measurement accuracy relative to reference angulations. Additional analyses examined the impact of input image count and model initialization on performance (see Supplementary Material S1 for more details). The fidelity of the 3D reconstruction was evaluated on both photographs of 3D-printed penile models and digitally rendered 3D CAD models. Quantitative assessment using geometric and image-based similarity metrics showed that reconstruction quality improved consistently with the number of input images, with optimal performance achieved using 12 standardized views. F-scores for printed model reconstructions increased from 0.59 with a single image to 0.82 with 12 images, accompanied by reductions in Chamfer distance and improvements across all perceptual metrics (Supplementary Table S1). Comparable trends were observed in the samples that the renders have been given as input to the 3D reconstruction model, where the mean F-score increased from 0.83 to 0.91 with additional views (Supplementary Table S2). The clinical utility of the pipeline was assessed by estimating curvature angles from reconstructed models of 182 digital single-curvature CAD samples, systematically varied in angulation, direction, location, and circumcision status (Supplementary Figure S4). Median angular error was lowest for mild curvatures (15°–45°) and increased with severity, reaching its highest values in cases with 105°–120° angulation. Directional effects were also observed: dorsal (upward) curvatures were most accurately measured (median error 4.8°, mean error 5.4°, 95% CI [3.8°, 7.0°]), whereas ventral (downward) curvatures yielded the highest error (median error 14.2°, mean error 14.4°, 95% CI [11.7°, 17.2°]), with left and right curvatures showing intermediate accuracy (median error 9.1°, mean error 12.7° 95% CI [9.1°, 16.3°], median error 11.5°, mean error 12.8°, 95% CI [10.0°, 15.5°], respectively). Curvature location along the shaft had minimal effect on accuracy, though slightly higher errors were observed in near-glans cases. Assessments on circumcised samples demonstrated modestly lower errors than uncircumcised samples. Across all samples, the median and mean difference between predicted and ground truth curvature values was 8.64° and 11.4° (95% CI [10.0°, 12.9°]), respectively. When comparing assessments on reconstructed models with those on original ground truth renders from identical viewpoints, the correlation coefficient was 0.902 (Supplementary Figure S5), and the correlation with original design specifications was 0.86, demonstrating strong agreement and validating the end-to-end accuracy of the pipeline from image input to curvature estimation. Processing time per sample was approximately 94 s on a workstation, with opportunities for further acceleration through parallelization. This work presents a fully automated, AI-driven pipeline for reconstructing 3D penile anatomy from multi-view photographs and performing objective PD assessment. Validated on a large, diverse set of synthetic and 3D-printed models, the method achieved high accuracy for clinically common curvatures while eliminating the need for specialized hardware or manual measurements. By enabling standardized, reproducible evaluation in both clinical and at-home settings, this approach could transform PD assessment and expand applications to treatment monitoring, surgical planning, and training. The pipeline is designed with transparency in mind, producing intermediate outputs throughout the reconstruction and assessment process, which can help clinicians better interpret results. Evaluation on synthetic and 3D-printed models demonstrated strong performance across a wide range of PD presentations, including variations in curvature angle, direction, location, and circumcision status. The custom CAD generation tool used in this study enables rapid creation of anatomically varied penile models, opening doors for the next generation of AI models training in this area of research. Compared with prior approaches that relied on small datasets, manual annotation, or hardware-intensive 3D scanning, our pipeline was validated on an unprecedentedly large set of models, providing a more comprehensive evaluation of automated 3D reconstruction and curvature analysis for PD. Despite these promising results, several limitations remain. Validation was conducted under controlled conditions using synthetic and printed models, and real clinical images will present additional challenges, such as lighting variation, background clutter, inconsistent orientation, and lack of a standardized method for erectile rigidity. Moreover, assessment currently relies on 2D projections rather than fully 3D-native analysis, and no global confidence measure is yet integrated to guide clinical decision-making. Another limitation is the number of images required for high-quality reconstruction; currently, 12 images yield optimal results. Capturing this many can be challenging in clinical settings; however, technologies like burst mode can help acquire them efficiently. Addressing these limitations will be essential for translation into routine practice. Future work includes validating the pipeline on a large, diverse cohort to ensure robustness, fine-tuning for extreme curvatures, enabling direct 3D-native analysis, and integrating confidence scoring to enhance reliability and clinical adoption. A.B.: Conceptualization-Equal, Data curation-Supporting, Formal analysis-Equal, Investigation-Equal, Methodology-Equal, Software-Equal, Visualization-Equal, Writing – original draft-Equal, writing – review & editing-Equal. E.J.: Data curation-Supporting. R.F.: Conceptualization-Equal, Data curation-Equal, Investigation-Equal, Project administration-Equal, Resources-Equal, Supervision-Equal, Writing – review & editing-Equal. F.H.: Conceptualization-Equal, Data curation-Equal, Investigation-Equal, Methodology-Equal, Funding acquisition-Lead, Project administration-Equal, Resources-Equal, Supervision-Equal, Writing – review & editing-Equal. This work was supported in part by the Canadian Urological Association Scholarship Foundation (CUASF) and the Canadian Urological Association (to F.H.), Natural Science and Engineering Council of Canada (NSERC) Discovery Grants (RGPIN-05144 to F.H.). Dr Ryan Flannigan is a consultant for Boston Scientific and Coloplast, has received speaking honoraria from Ferring, holds a fellowship educational grant from Boston Scientific, and has received research and educational grants from TerSera. He is also a co-founder, shareholder, and director of Teumo Health Technologies. Dr Faraz Hach is a co-founder, shareholder, and director of Teumo Health Technologies.

Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.