Inspiration
We were inspired by the challenge of extracting useful anatomical information from CT scans without relying on pre-labeled datasets or heavy deep-learning models. We wanted to see how far we could get using classical image-processing techniques and the information already provided by the aorta mask.
What We Learned
We jokingly say we learned a lot about biology, but we also learned a surprising amount about image processing, kernel-based operations, and the complexities of working with medical image modalities like CT. Small differences in intensity, geometry, and preprocessing could have a major impact on the final result.
How We Built It
We developed a CPU-based pipeline that starts from the provided aorta segmentation and searches the surrounding CT volume for potential daughter vessels. We experimented with intensity thresholds, morphological operations, connected components, vesselness, geometric filtering, and vessel tracking to identify candidate arteries and estimate their ostia, seed points, directions, and radii.
Challenges
One of our biggest challenges was uncertainty about correctness: without ground-truth labels, it was difficult to know whether a detection was genuinely correct. We also struggled with reliably getting detected vessels to connect back to the aorta, especially when intensity or image quality varied. Finally, balancing precision and recall was an ongoing challenge—making the detector more sensitive often introduced false positives, while stricter filtering caused obvious vessels to be missed.
Log in or sign up for Devpost to join the conversation.