E3278. Automated Comprehensive Fat Quantification and Sarcopenia Assessment Tool: A Potential MRI Biomarker
Authors
Jordan Sim;
Tan Tock Seng Hospital
Arvind Srinivasa;
Agency for Science, Technology and Research (A*STAR)
Ling Yun Yeow;
Agency for Science, Technology and Research (A*STAR)
Audrey Yeo;
Tan Tock Seng Hospital
Wee Shiong Lim;
Lee Kong Chian School of Medicine, Nanyang Technological University; Tan Tock Seng Hospital
Bhanu Prakash;
Agency for Science, Technology and Research (A*STAR)
Cher Heng Tan;
Lee Kong Chian School of Medicine, Nanyang Technological University; Tan Tock Seng Hospital
Objective:
Body composition data from abdominal MRI studies has the potential to identify patients at risk of metabolic syndrome and sarcopenic obesity, amongst other chronic conditions. Chronic diseases require long-term management and lead to disability, reduced quality of life, and increased healthcare expenditure. Screening and early detection are the cornerstone of early intervention. This study aims to develop a client-server cloud-based application suite capable of end-to-end quantitative and qualitative analysis of abdominal fat/muscle and visceral adiposity.
Materials and Methods:
A total of 190 community-dwelling older adults were evaluated (mean age, 68 ± 8 years; 69.5% women; mean BMI, 23.75 ± 3.65 kg/m<sup>2</sup>). A MultiRes Attention U-Net with hybrid loss function was proposed for adipose tissue segmentation. MRI scans from 26 participants were manually segmented to generate ground truth. Data augmentation was performed using MR data acquisition variations. The model output can be used for construction of sarcopenia prediction model, fatty liver quantification, and pancreatic fat assessment. In compliance with the hospital’s security standard, the tool was deployed as a client-server cloud-based application for metabolic quantification and sarcopenia prediction.
Results:
For abdominal fat segmentation, the median Dice scores were 0.97 for superficial subcutaneous adipose tissue (SSAT) and deep subcutaneous adipose tissue (DSAT), and 0.96 for visceral adipose tissue (VAT). Mean Hausdorff distance is <5 mm for all the three abdominal fat compartments. For upper thigh muscle/fat segmentation, the median Dice scores are 0.58 for intermuscular adipose tissue (IMAT), 0.92 for SSAT, and 0.88 for muscle. Comparatively lower scores for IMAT could be due to small and heterogeneous distribution of fat, offering an opportunity for refinement of the algorithm.
Conclusion:
An accurate, automated, and comprehensive fat quantification and sarcopenia assessment tool with reproducible results holds promise for body-profiling in large cohort studies.
