FatSegNet: A fully automated deep learning pipeline for adipose tissue segmentation on abdominal dixon MRI

Abstract

Purpose: Introduce and validate a novel, fast, and fully automated deep learning pipeline (FatSegNet) to accurately identify, segment, and quantify visceral and subcutaneous adipose tissue (VAT and SAT) within a consistent, anatomically defined abdominal region on Dixon MRI scans. Method: FatSegNet is composed of three stages: (i) Consistent localization of the abdominal region using two 2D-Competitive Dense Fully Convolutional Networks (CDFNet), (ii) Segmentation of adipose tissue on three views by independent CDFNets, and (iii) View-aggregation. FatSegNet is validated by: 1) comparison of segmentation accuracy (sixfold cross-validation), 2) test-retest reliability, 3) generalizability to randomly selected manually re-edited cases, and 4) replication of age and sex effects in the Rhineland Studya large prospective population cohort. Results: The CDFNet demonstrates increased accuracy and robustness compared to traditional deep learning networks. FatSegNet Dice score outperforms manual raters on VAT (0.850 vs. 0.788),and produces comparable results on SAT (0.975 vs. 0.982). The pipeline has excellent agreement for both test-retest (ICC VAT 0.998 and SAT 0.996) and manual re-editing (ICC VAT 0.999 and SAT 0.999). 2 SUBMITTED TO MAGNETIC RESONANCE IN MEDICINE Conclusion: FatSegNet generalizes well to different body shapes, sensitively replicates known VAT and SAT volume effects in a large cohort study, and permits localized analysis of fat compartments. Furthermore, it can reliably analyze a 3D Dixon MRI in ∼ 1 min, providing an efficient and validated pipeline for abdominal adipose tissue analysis in the Rhineland Study. K E Y W O R D S Subcutaneous adipose tissue, Visceral adipose tissue, Dixon MRI, Neural networks, Deep learning, Semantic segmentation 1 | INTRODUCTION The excess of body fat depots is an increasing major public health issue worldwide and an important risk factor for the development of metabolic disorders and reduced quality of life [1, 2]. While the body mass index (BMI) is a widely used indicator of adipose tissue accumulation in the body, it does not provide information on fat distribution [3] -neither with respect to different fat tissue types nor with respect to deposit location. Different compartments of adipose tissue are associated with different physiopathological effects [4, 5]. Abdominal adipose tissue (AAT), composed of subcutaneous and visceral adipose tissue (SAT and VAT), has long been associated with an increased risk of chronic cardiovascular diseases, glucose impairment, and dyslipidemia [6, 7]. Recently, several studies have indicated a stronger relation between the accumulation of VAT with an adverse metabolic and inflammatory profile compared to SAT [8, 9]. Therefore, an accurate and independent measurement of VAT and SAT volumes (VAT-V and SAT-V) is of significant clinical and research interest. Currently, the gold standard for measuring VAT-V and SAT-V is the manual segmentation of abdominal fat images from Dixon magnetic resonance (MR) scans -a very expensive and time-consu...

FastSurfer - A fast and accurate deep learning based neuroimaging pipeline

“…With the exception of our recent work in ( Estrada et al, 2018 , 2019 ), dense connections within convolutional blocks have been implemented via concatenation of feature maps (see QuickNAT ( Roy et al, 2019 )) - effectively doubling the numbers of learnable parameters ( Fig. 3 top) within each encoder and decoder block and thus considerably increasing memory requirements.…”

“…3 top) within each encoder and decoder block and thus considerably increasing memory requirements. Here, we employ competitive dense blocks in which concatenations are replaced with maxout activations ( Goodfellow et al, 2013 ; Estrada et al, 2018 , 2019 ). The maxout activations induce competition between feature maps and significantly reduce the number of parameters compared to the classical dense blocks, thus creating a lightweight model ( Fig.…”

“…Each F–CNN has the same encoder/decoder-based architecture with skip connections ( Ronneberger et al, 2015 ), enhanced with unpooling layers ( Noh et al, 2015 ) and dense connections ( Huang et al, 2017 ) within each block. The main methodological improvements of FastSurferCNN compared to QuickNAT are the introduction of competition within each block (competitive dense blocks) by replacing concatenation with maxout operations ( Goodfellow et al, 2013 ; Estrada et al, 2018 , 2019 ), as well as the inclusion of a wider image context within each 2D F–CNN (spatial information aggregation) in order to retain enough spatial information for the accurate segmentation of neuroanatomical structures, such as cortical gray matter regions.…”

Large-scale biometry with interpretable neural network regression on UK Biobank body MRI

“…Various methods have been proposed for the delineation of individual adipose tissue depots in these images 12 . Among other techniques, automated image analysis with convolutional neural networks for segmentation has become an established technique for images of this kind 13,14 as well as for CT images 15,16 . However, these systems learn to perform segmentation from training data in the form of reference segmentations, which must accordingly be carefully prepared, often with substantial amounts of manual guidance.…”

Fetal Cortical Plate Segmentation Using Fully Convolutional Networks With Multiple Plane Aggregation

“…In addition, the multi-view aggregation technique is used to enhance the segmentation accuracy by applying a 3D information to a 2D deep learning network. It combines three results from separate learning networks of 2D slices from three orthogonal planes (axial, coronal, and sagittal) to generate the final segmentation ( Guha Roy et al, 2019 ; Jog et al, 2019 ; Estrada et al, 2020 ). The test-time augmentation (TTA) technique can obtain more robust prediction results using multiple predictions for a single input by applying the augmentation to test data, which is often used for the training phase in deep learning networks ( Matsunaga et al, 2017 ; Jin et al, 2018 ).…”