TOWARDS PRECISION OSTEOPOROSIS SCREENING: A HOLISTIC DEEP LEARNING APPROACH TO INTEGRATING IMAGING, PATIENT DATA, AND LONGITUDINAL INSIGHTS

Author

• Harikrishnan Changarnkothapeecherikkal, Michigan Technological UniversityFollow

Date of Award

2026

Document Type

Open Access Dissertation

Degree Name

Doctor of Philosophy in Computational Science and Engineering (PhD)

Administrative Home Department

Department of Applied Computing

Advisor 1

Guy Hembroff

Committee Member 1

Jeffrey Wall

Committee Member 2

Chad L. Klochko

Committee Member 3

Donald Peck

Abstract

Osteoporosis is a silent yet devastating bone disease that affects millions worldwide, often remaining undiagnosed until a fragility fracture occurs. While dual-energy X-ray absorptiometry (DXA) remains the gold standard for diagnosis, its limited availability and stringent screening thresholds have led to a critical gap in early detection. This dissertation presents a comprehensive approach to address this challenge by leveraging deep learning for opportunistic osteoporosis screening using routine knee radiographs.

The research develops a multimodal deep learning framework that integrates Posterior-Anterior (Frontal) and Lateral knee radiographs with patient-specific clinical covariates to identify individuals requiring confirmatory DXA scanning. The methodology begins with the automated extraction and preprocessing of a robust dataset from clinical PACS environments, followed by the application of targeted image enhancement techniques. A multimodal fusion model is then designed and validated to accurately assess osteoporosis risk by mathematically combining these enhanced visual features with demographic tabular data.

To ensure algorithmic equity and clinical deployability, this work introduces FairOsteoFusionNet, an architecture that systematically mitigates intersectional demographic bias. By employing a dynamic gating mechanism, the network autonomously down-weight corrupted radiographic signals caused by adipose-induced soft-tissue scattering in high-BMI cohorts, shifting reliance toward clinical risk profiles to maintain equitable performance. Furthermore, the dissertation develops a verifiable semantic auditing pipeline. By integrating YOLO-based anatomical segmentation and illumination contrast ratios, the framework mathematically quantifies the primary vision model's spatial attention. A localized Large Language Model, aligned via Direct Preference Optimization (DPO), translates these objective metrics into causal, expert-level clinical narratives that transparently justify the diagnostic predictions based on physical biological markers.

Finally, the research outlines a proposed temporal analysis framework to monitor longitudinal changes in bone quality and forecast future fracture risk. Standing at the intersection of computer vision and medical imaging, this work provides a complementary, highly interpretable triage tool. By transforming routine imaging into an equitable screening mechanism, this dissertation aims to facilitate earlier intervention, reduce fracture incidence, and improve accessible healthcare delivery.

Recommended Citation

Changarnkothapeecherikkal, Harikrishnan, "TOWARDS PRECISION OSTEOPOROSIS SCREENING: A HOLISTIC DEEP LEARNING APPROACH TO INTEGRATING IMAGING, PATIENT DATA, AND LONGITUDINAL INSIGHTS", Open Access Dissertation, Michigan Technological University, 2026.

https://digitalcommons.mtu.edu/etdr/2153