Repository logo
 

An Adapted Convolutional Neural Network for Automatic Measurement of Pancreatic Fat and Pancreatic Volume in Clinical Multi-Protocol Magnetic Resonance Images: A Retrospective Study With Multi-Ethnic External Validation of Various Fat Deposition

Files

Thesis(3.64 MB)

Date

2022

Authors

Supervisor

Lu, Jun
Murphy, Rinki

Item type

Thesis

Degree name

Doctor of Philosophy

Journal Title

Journal ISSN

Volume Title

Publisher

Auckland University of Technology

Abstract

Anthropometric indices, such as body mass index (BMI), waist circumference (WC), and waist to height ratio (WHtR) have limitations in accurately predict the pathophysiology of diabetes mellitus, cardiovascular diseases, and metabolic syndrome due to ethnic differences in fat distribution. Recent studies showed that the visceral adipose tissue (VAT) deposition and fat content of internal organs, most notably intra-hepatic and intra-pancreatic fat, has emerged as a more important parameter. Measurement of fat fraction is now regarded as a challenge in clinical settings. Magnetic Resonance Imaging (MRI) based quantification of fat fraction requires highly accurate data reconstruction for the assessment of hepatic and pancreatic fat accumulation in medical diagnostics and biomedical research. So, automated pancreas segmentation and accurate fat content determination from medical images are important for clinical and research applications, including type 2 diabetes risk prediction. In this study, we tried to assess the coordination between the traditional anthropometric indices and the various fat depositions within different ethnicities in New Zealand. We further established the signal model of oil and water emulsion used for phantom study with a field strength of 3.0 T. Finally, we modified a novel convolutional neural network (CNN) for pancreatic volume and fat fraction segmentation in magnetic resonance imaging (MRI) scans. We used 104 previous participants with different ethnic backgrounds, including New Zealand European, Māori (the indigenous people of New Zealand), Pacific Islanders (PI), and Asian. Their weight, height, and waist circumference (WC) were measured, and subcutaneous, visceral, intra-hepatic, and intra-pancreatic fat depositions were obtained using magnetic resonance imaging (MRI). The results showed VAT depositions, but not subcutaneous adipose tissue (SAT) depositions, varied significantly at all levels among the three groups. BMI was best associated with L23SAT in NZ Europeans (30%) and L45VAT in Māori/PI (24.3%). Overall, WC and WHtR were correlated well with L45SAT (18.8% and 12.2% respectively). Intra-pancreatic fat deposition had positive Pearson relation with NZ European’s BMI and Māori/PI’s WC, but no regression correlation with anthropometric indices. Conventional anthropometric indices do not correspond to the same fat depositions across different ethnic groups. To accomplish the phantom study for machine learning, we used fat fraction quantification from phantom as a standard gradient to compare hepatic fat fraction and intrapancreatic fat fraction quantification with both algorithm of magnetic resonance spectroscopy (MRS) and Iterative Decomposition with Echo Asymmetry and Least-squares estimation (IDEAL) in MRI. We also compared MRS and IDEAL pancreatic fat fraction quantification with expert manual pancreatic fat measurement. We noticed a strong correlation between true fat volume fraction and the fat fractions from both IDEAL (R2=0.99) and MRS (R2=0.99). Linear correlation and Pearson’s correlation were applied to both the phantom and in vivo measurements. The results of in vivo measurement demonstrated a good correlation between MRI measurements of hepatic fat fractions, but varied for the pancreatic fat fraction. We also observed that the manual operation performed better than IDEAL and MRS pancreatic fat fractions reading, which helped with the further establishment of auto pancreatic fat measurement by machine learning. In this retrospective, prognostic study, we conducted pancreatic boundary identification and fat fraction segmentation. Images were modified to allow CNN with improved super pixel pre-processing. The formal training and testing of the artificial intelligence was established on more than 3,000 abdominal MR images. Validation was then conducted on 200 images from 10 additional patients who were each scanned twice. Our algorithm achieved a dice similarity coefficient (DSC) of 91.2%. This is the first algorithm for automated pancreas volume and intra-pancreatic fat determination with > 90% DSC, which has the potential to be widely used for rapid and accurate pancreatic fat quantification in research and clinical settings when using abdominal MRI.

Description

Keywords

Source

DOI

Publisher's version

Rights statement

Permanent link

https://hdl.handle.net/10292/15315

Collections

Doctoral Theses