PII-184 - TOWARDS ASSISTING POTENTIAL PREVENTION CLINICAL TRIALS OF BONE-PROTECTIVE THERAPIES FOR DUCHENNE MUSCULAR DYSTROPHY USING DEEP SEGMENTED IMAGING BIOMARKER

S. Kang¹, R. Ravindranunni², O. Elashkar³, H. Ju⁴, R. Willcocks⁵, S. Kim⁶; ¹University of Florida College of Pharmacy Department of Pharmaceutics, ²University of Florida Department of Physical Therapy, ³Department of Pharmaceutics, College of Pharmacy, University of Florida, ⁴NVIDIA, ⁵Department of Physical Therapy, University of Florida, ⁶Center for Pharmacometrics and Systems Pharmacology, College of Pharmacy, University of Florida.

Background: Bone fragility is one of the major complications associated with Duchenne muscular dystrophy (DMD). Corticosteroids, often given to individuals with DMD as supportive care, worsen the cortical bone condition. Quantitative measures of bone fragility are critical to monitor longitudinal trajectories over DMD progression. This study aims to quantify longitudinal changes in bone cortical thickness over DMD progression by utilizing advanced deep-learning driven segmentation techniques using magnetic resonance imaging (MRI) data.

Methods: The MRI dataset utilized in our study comprises longitudinal DICOM images and was collected from 56 boys across multiple sites over 13 years through natural history ImagingDMD study (NCT01484678). Labels were generated manually using 3D Slicer to ensure the precise training of the model. For auto-segmentation, the Dynamic UNET was employed as our multi-segmentation model and implemented through MONAI Label and 3D Slicer that assist medical image analysis with deep-learning framework. We trained the model with an evaluation metric Dice loss and fine-tuned hyperparameters to identify the optimal model. With the trained segmentation model, we computed cross-sectional area of the cortical bone from the mid-femur region using deep-segmented MRI images that were unused at training. HiPerGator, the University of Florida supercomputer, was utilized.

Results: A total of 100 images were annotated and subsequently divided into separate training, validation and test dataset. The trained segmentation model demonstrates an accuracy of 97.49% in segmenting the 3-classes. We can get the cross-sectional area of bone and marrow on each slice based on the segmentation (Figure 1), and employ the averaged cortical thickness as a biomarker.

Conclusion: The developed automated deep learning-based imaging analysis tool will accelerate cortical bone analysis by accurately segmenting and calculating cross-sectional areas, which will ultimately facilitate the identification of biomarkers for prevention clinical trials examining bisphosphonates and other bone-protective therapies. Through the quantification of the longitudinal changes in the deterioration of cortical bone thickness, clinicians can intervene and mitigate potential side effects at the optimal moments.