Tracing and Segmentation of Penile Vessels from Computed Tomography Angiograms
Abstract
Purpose
Penile vessels’ morphology and integrity are highly relevant to diagnosis and management of erectile dysfunction, as well as treatment planning for sexual function-sparing pelvic radiotherapy. Automatic segmentation of penile vessels on CT angiograms is challenging due to the difficulty of manual labeling, small vessel diameters, and confounding appearances of high contrast calcifications. We develop a model-driven approach incorporating geometric priors on vessel configurations.
Methods
We leverage knowledge that major penile vessels run along the shaft. After initial shaft segmentation, we extract high contrast voxels as noisy point samples of vessel paths and construct a graph with direction-dependent edge energies. Corresponding proximal and distal endpoints are estimated using Kalman filtering along the shaft direction, with Mahalanobis gating for identifying early terminations and testing false point samples. The vessel tracing problem is mapped to a multi-agent pathfinding optimization setting by equating the shaft arc-length parameter with agent timing. The resulting vessel trace is rasterized and an active contour model is used to refine the final vessel estimate. Evaluation was performed on segmenting 30 vessels from 16 CTAs. Vessel curvature and cross-sectional area were used to assess anatomical realism.
Results
Vessel curvatures exhibited a moderate to strong correlation with shaft curvature, with median, IQR, and range of 0.658/[0.556, 0.771]/[0.254, 0.928], consistent with the anatomical expectation of alignment between dorsal vessels and the shaft. Cross-sectional areas from our model (median/IQR/range: 10.9/[6.73, 16.5]/[0.391, 32.1]) suggest potential oversegmentation compared to our expectation of ~2 mm2 based on vessel diameters reported in literature.
Conclusion
A model-driven method was developed for segmenting fine penile vessels. Further development is necessary and underway for more extensive validation on larger datasets and parameter sensitivity analysis.