Deep learning model segments penile tissue in 34K+ MRI scans

A team led by Jan Ernsting developed a deep learning framework for whole-penis segmentation in multi-channel DIXON MRI. Using a curated dataset of 145 subjects (13,050 annotated slices) and a double-annotated test set of 24 subjects (2,160 slices), they optimized a 3D nnU-Net architecture. The model achieved a 5-fold cross-validation Dice score of 0.90 and an independent test Dice of 0.92 with Hausdorff distance of 3.58, matching expert-level accuracy.

Deployed on 34,412 UK Biobank participants, the framework enables automated quantification of total penile tissue — both external and internal components. Longitudinal evaluation in 2,282 men showed high reproducibility (r=0.87). The authors will publicly release trained model weights, providing an open resource for urological imaging studies and quantitative phenotyping of male reproductive health conditions like micropenis and endocrine disorders.