External Validation of an Automated Segmentation Tool for Abdominal Fat Tissue using DIXON MRI: Data from the CutDM trial

Background and rationale

Segmentation and quantifying the volume of visceral adipose tissue (VAT) and subcutaneous adipose tissue (SAT) is an emerging field that has shown to be highly relevant for studying the risk of various metabolic diseases and monitoring treatment effects over time. The reference method for segmenting VAT and SAT involves manual segmentation in a prespecified volume, such as delimited by the cephalic part of L3 to the caudal part of L5 of the lumbar spine using the fat image produced by the DIXON technique. However, scalability and more widespread use is challenged by the time-consuming nature and potential inter- and intrareader variability of the manual approach.

Lately increasing number of segmentation tools have been made available under open source such as the FatSegNet. For such tools, only the internal diagnostic test accuracy is reported. Furthermore, segmentation tools rarely have a proper localizer which is required for an end-to-end workflow.

Objectives

We will prepend the existing FatSegNet model with a L3 to L5 delimiter model to establish a pipeline for fully automated analysis of an axial Dixon stack. The aim of this study will be to test the diagnostic accuracy this full VAT-SAT quantification approach.

Methods

Two experienced readers will manually segment the VAT and the SAT for each slice of the abdominal volume defined by the axial slice through the cephalic part of the L3 vertebra to caudal part of the L5 vertebra in 32 participants (16 male and 16 female) with BMI >25 who underwent 6-point mDIXON QUANT (Philips Healthcare) at baseline and follow up.

Population

The sample for this study inherits the characteristics of the CUTDm trial with the eligibility criteria:

• Diabetes mellitus type 2

• BMI > 25

• Men or postmenopausal women aged 18-75 years. Menopause will be defined as >12 months without menses

• HbA1c 48-75 mmol/mol (6.5%-9.0%)

• Treated with or without Metformin, DPP-4 inhibitors, SGLT-2 inhibitors and/or GLP-1 receptor agonists (GLP-1RA)

• Non-smokers or having quitted smoking >1 year before inclusion in the study

• Acceptance of regulation of antidiabetic, antihypertensive, and lipid-lowering medications by the Cut-DM endocrinologists only

Index test

The segmented volumes for VAT and SAT outputted by the AI tool.

Reference test

The scan-level mean segmented volumes of the two independent, reference readers.