SEHO YUK, MS Candidate: No financial relationships to disclose
Background: Osteoporosis is a chronic condition marked by weakened bone structure and mineral loss, particularly affecting postmenopausal women. The current diagnostic gold standard, dual-energy X-ray absorptiometry (DXA), is subject to significant variability arising from clinician experience, human error, and environmental conditions. While machine learning (ML) offers potential solutions, achieving high accuracy often requires massive datasets that are logistically difficult to acquire. This study aims to overcome these limitations by developing a diagnostic model that integrates a K-Nearest Neighbor (kNN) algorithm with a custom honeycomb structure analysis library to quantitatively analyze bone porosity using a smaller dataset.
Materials and
Methods: We retrospectively collected 300 de-identified DXA images from a patient cohort (2000–2023), categorized into normal, osteopenic, and osteoporotic groups based on T-scores. The analytical workflow involved rigorous image preprocessing, including contrast enhancement (equalization) and binarization, to isolate vertebral bone segments and define inner and outer boundaries. A custom honeycomb structure library was applied to these preprocessed images to calculate quantitative features such as pore size in arbitrary units (AU) and percentage of bone pore space. The model was trained to identify the minimum data volume required to achieve stability and diagnostic accuracy.
Results: Quantitative analysis revealed significant distinctions in bone architecture across the three conditions. Normal bone exhibited an average pore size of 77.19 AU and occupied 4.04% of the pore space. In contrast, osteopenic bone measured 264.33 AU (11.29% space), and osteoporotic bone measured 279.54 AU (12.70% space). The mean pore size in osteopenic and osteoporotic bones was approximately 2.63 and 2.95 times larger than in normal bone, respectively. The combined imaging and quantitative analysis achieved model stability and training accuracy with fewer than 300 pre-treated images per condition, contradicting the assumption that massive datasets are strictly necessary for effective ML modeling in this domain.
Conclusion: This study demonstrates that quantified bone pore space is a reliable, objective metric for assessing skeletal health. The findings suggest that extensive, resource-intensive datasets are not required if images are rigorously preprocessed to highlight key features. The proposed honeycomb structure analysis offers a robust, less subjective, and efficient alternative to traditional visual DXA assessments for diagnosing osteoporosis.
*Unless otherwise noted, all abstracts presented at ENDO must not be released to the press or the public until the date and time of presentation. For oral presentations, the abstracts are embargoed until the session begins. The Endocrine Society reserves the right to lift the embargo on specific abstracts that are selected for promotion prior to or during ENDO.*
.jpg)
.jpg)
