Project
Ultrasound Bone Age Assessment
Capstone Research Project
Electrical and Computer Engineering, University of Alberta
Jan 2026 - Apr 2026
Edmonton, Alberta, Canada · On-site
Capstone research project at the University of Alberta focused on building a fully automated ultrasound-based bone age assessment pipeline. The system combined long-sequence transformers, segmentation, and clinical staging models to process pediatric ultrasound video and deliver radiology-grade skeletal maturity signals.
Worked under professor Dr. Edmond Lou and Dr. Marek Reformat.
Developed the first fully automated pipeline for extracting skeletal maturity features from 2D ultrasound video, removing the need for ionizing radiation in pediatric bone age assessment.
Designed a transformer framework for ultra-long medical video sequences of 900 to 5,828 frames with full self-attention, achieving 91% keyframe localization accuracy across 17 bone landmarks.
Validated the same architecture across ultrasound, cardiac echocardiography, and laparoscopic surgery without architectural changes, demonstrating domain-agnostic generalizability.
Adapted Mask2Former for fragmented bone segmentation in ultrasound, improving epiphysis detection from 11.9% with YOLOv8 to 83.1% mAP@50.
Achieved 97.85% Risser sign classification accuracy within ±1 stage on held-out patient data, matching radiologist inter-observer agreement levels.
Deployed the full system as a Dockerized desktop application processing videos in about 20 seconds on GPU.
Medical AITransformersSegmentation