PT-018 - AI/ML DERIVED MAGNETIC RESONANCE BIOMARKERS FOR BONE FRAGILITY IN DUCHENNE MUSCULAR DYSTROPHY

S. Kang¹, H. Jeong², R. Kunnath Ravindranunni², H. Ju³, G. Walter², K. Vandenborne², R. Willcocks², S. Kim²; ¹University of Florida College of Pharmacy Department of Pharmaceutics, Gainesville, FL, USA, ²University of Florida, Gainesville, FL, USA, ³NVIDIA, CA, USA.

Background: Duchenne muscular dystrophy (DMD) is a rare pediatric disease, causing progressive muscle weakness and functional disability. Bone fragility is one of the major complications associated with DMD. Quantitative magnetic resonance (qMR) muscle imaging biomarkers have shown strong correlations with functional endpoints that have been widely used in DMD clinical trials. This study aimed to extract additional qMR biomarkers that are clinically meaningful to quantify longitudinal changes in bone fragility over DMD progression through a deep segmentation model and voxel-wise feature extractions.

Methods: The qMR dataset utilized in our study includes longitudinal data collected from 13 individuals with DMD (NCT01484678), each scan comprised 3 slices at the mid-femur. The cortical bone and marrow areas of each qMR were segmented using a trained DynUNET. PyRadiomics enabled the extraction of a broad range of features, focusing on shape features. Distance transformation using the Euclidean distances enabled the calculation of the thickness of cortical bone. Top shape features were selected based on the Pearson correlation coefficients, which compared correlations between imaging features extracted using PyRadiomics and MR spectroscopy fat fraction measures from vastus lateralis (FFVL) muscle as an index of disease severity. In addition, subgroup analysis was performed to investigate the changes in the shape features extracted from cortical bone and marrow.

Results: The trained segmentation model performed well with Dice Score of 97.49%, which is a metric for assessing segmentation accuracy. As demonstrated in Figure, all the shape features and cortical thickness extracted from bone images gradually changed as DMD progressed, as shown in FFVL. Notably, subgroup analysis within the 8 individuals revealed a significant difference in cortical thickness among steroid users, with a p-value of 0.057 from the t-test.

Conclusion: This study demonstrates that AI-assisted radiomic feature extraction effectively identifies key features from MRI data, with the extracted features serving as qMR biomarkers for assessing bone fragility. Furthermore, the features of cortical bone correlated with FFVL, indicating that they capture worsening bone health with DMD disease progression.