Meeting Banner
Abstract #2824

Deep-learning based 3D segmentation of thigh muscle and classification of intramuscular fat on T1-weighted axial MRI

Upasana Upadhyay Bharadwaj1, Amir M. Pirmoazen1, Zehra Akkaya1, John A. Lynch2, Gabby B. Joseph1, Sharmila Majumdar1, Valentina Pedoia1, and Thomas M. Link1
1Radiology and Biomedical Imaging, University of California, San Francisco, San Francisco, CA, United States, 2Epidemiology and Biostatistics, University of California, San Francisco, San Francisco, CA, United States

Synopsis

Intramuscular fat is an important biomarker for knee osteoarthritis. Quantitative analysis on routine clinical imaging (T1-weighted MRI) is not feasible without pixel-level annotation, leading to the adoption of Goutallier classification, a semi-quantitative grading system that is time-consuming and has variable reproducibility. This study automates binarized Goutallier classification on patients (n=50) from the Osteoarthritis Initiative cohort with a two-staged process: deep-learning 3D segmentation of quadriceps and hamstrings (dice scores of 0.89[0.88,0.90] and 0.84[0.83,0.87], respectively) followed by histogram features for classification of intramuscular fat (0.93[0.92,0.95] AUROC). With model-reader kappa (0.64[0.61,0.68]) comparable to inter-reader kappa (0.61[0.59,0.64]), our approach shows promise for end-to-end automation.

This abstract and the presentation materials are available to members only; a login is required.

Join Here