AAI_2025_Capstone_Chronicles_Combined

1 Cervical Spine Fracture Detection Using Computer Vision AAI-590 Capstone Project (Team 4) ​

Andy Malinsky Master of Science in Applied Artificial Intelligence Shiley Marcos School of Engineering / University of San Diego​ amalinsky@sandiego.edu

Christopher Alleyne Master of Science in Applied Artificial Intelligence

Devin Eror Master of Science in Applied Artificial Intelligence Shiley Marcos School of Engineering / University of San Diego​ deror@sandiego.edu

Jory Hamilton Master of Science in Applied Artificial Intelligence

Shiley Marcos School of Engineering / University of San Diego​ calleyne@sandiego.edu

Shiley Marcos School of Engineering / University of San Diego joryhamilton@sandiego.e du

localization and were fine-tuned through threshold optimization and early stopping. Results showed that the CNN struggled with fracture detection, Faster R-CNN achieved moderate precision and recall, and DETR demonstrated high sensitivity with improved bounding-box localization. These findings suggest that deep learning models have potential to aid cervical fracture detection, though additional optimization and larger datasets are needed for clinical reliability. KEYWORDS Cervical Fracture Detection; Object Detection; Classification; Deep Learning; DETR; Faster R-CNN; Convolutional Neural Networks (CNN); Imbalanced Data; Medical Imaging; Diagnosis 1 Introduction Detecting bone fractures is critical in post-operative and trauma care because delayed or missed diagnoses can lead to paralysis, chronic pain, and long-term neurological complications. Traditional methods rely on radiologists and surgeons to manually interpret CT or MRI scans,

ABSTRACT

Detecting cervical spine fractures in CT imaging is challenging due to subtle fracture patterns, class imbalance, and variability in scan quality, and missed diagnoses can lead to severe neurological consequences. Using a curated subset of the RSNA 2022 Cervical Spine Fracture Detection dataset, this project explores whether deep learning-based computer vision models can support clinicians by automating fracture identification and localizing suspected fracture regions. We developed a preprocessing and modeling pipeline that included data normalization, undersampling, bounding-box validation, and a patient-level stratified split to prevent data leakage. Three models were implemented and evaluated: a baseline Convolutional Neural Network (CNN), a Faster R-CNN object detector, and DETR transformer-based detector. The CNN established a baseline for binary classification performance, while Faster R-CNN and DETR enabled spatial

304