FP6.7 Constructing Large-Scale 3D Hip Morphology Database of Normal Hip Development Using Multi-Modal Artificial Intelligence Tools

Introduction: Hip morphology has significant role in a range of hip pathologies. While some of these anatomical features can be readily measured from plane films, the more complex and clinically significant dysmorphologies require 3D evaluations (e.g., from CT or MRI). Comprehensive 3D evaluation of hip morphology is cumbersome and often neglected clinically. Moreover, there is a lack of population-specific thresholds for those assessments as the existing data is based on 2D measurements from X-Ray images and are not directly translatable to 3D assessments. Here we aimed to use multi-modal AI pipelines to generate a comprehensive registry of normal hip development during skeletal development and maturation form a large dataset of clinical CT scans at a tertiary-care children’s hospital.

Methods: Following IRB approval, we identified all the abdominal/pelvic/hip scans done at our institute between 2012-2020. We then developed a multi-modal pipeline to automatically: 1) identify any documented hip pathologies in the radiology reports using natural language processing (NLP), 2) reconstruct 3D models of the hip bones and identify anatomical landmarks using convolutional neural networks and deep learning, and 3) to measure hip anatomy in 3D using a custom and validated automatic software.

Results: The NLP pipeline achieved an accuracy of 0.98 in identifying hip pathology from radiology reports. The 3D reconstruction and landmark detection pipeline resulted in average Dice coefficient of 0.98±0.03 and average surface error of <1 mm. The morphology measurement pipeline resulted in an average error of <2 mm and <6 degrees. From a total of 52,360 CT scans, we identified and analyzed 9,721 “good quality” normal CT scans (49.3% Females; Age: 7 to 25 years, average: 14 4 years; 19,442 hips). On average, females had smaller femoral heads, epiphyseal tubercle, femoral necks, acetabulum, and alpha angles along with greater peripheral cupping, coronal head-neck tilt, femoral head-neck offset, overall femoral head coverage, acetabular anteversion, and posterior-superior center-edge angles (P<0.001). There were no clinically meaningful sex-differences in anterior-superior center-edge angles and sacro-pelvic sagittal alignments.

Summary/Clinical Significance: The current project highlights the feasibility of multi-modal approaches to process existing clinical data to generate large-scale registries, which can then be used to improve care through evidence-based personalized diagnosis and treatment planning. This rich database is currently being used to develop normative growth charts for detailed anatomical features of the hip throughout the skeletal growth and maturation. We are planning to publicly release this data to assist with personalized assessment of hip dysmorphology.