Automated segmentation of the human supraclavicular fat depot via deep neural network in water-fat separated magnetic resonance images

Figures

• 

  Figure 1. Comparison between raw manual annotation and processed labels. (A) Axial; (B) coronal. The supraclavicular fat depots are shown in red. As shown, most non-fatty tissues are removed from the supraclavicular fat depot mask after label pre-processing.

• 

  Figure 2. An overview of the whole network. It consists of three combining 2D U-Net-like networks and a 3D fusion network to mimic the workflow of physicians for characterizing BAT regions and to efficiently encode the multi-modal information and extract the 3D context information from multi-modal MRI scans for the segmentation of the BAT. The three combining 2D networks leverage multi-modal information and comprehensive 2D context information in axial, coronal, and sagittal planes to conduct the preliminary segmentation and the 3D fusion network combines multi-modal information, 3D context information and preliminary segmentation results for obtaining a fine-tuning segmentation. 2D, two-dimensional; 3D, three-dimensional; BAT, brown adipose tissue; MRI, magnetic resonance imaging.

• 

  Figure 3. The detailed architecture of the 3D fusion network. The 3D fusion network is a shallow 3D neural network designed to synthesize multi-modal MRI scans and preliminary segmentation results obtained from three combining 2D networks for obtaining a fine-tuning segmentation by leveraging 3D context information. The 3D fusion network consists of four 3D convolution modules and a final 3D 1×1×1 convolution layer with a softmax activation function to map 3D features to the final segmentation prediction. Each 3D convolution module has a 3×3×3 convolution layer with the batch normalization and leaky ReLU activation function. 2D, two-dimensional; 3D, three-dimensional; ReLU, rectified linear unit; MRI, magnetic resonance imaging.

• 

  Figure 4. The illustration of the pre-processing pipelines for MRI scans (PDFF, T2*) and manual annotation labels. PDFF, proton density fat fraction; F, fat; W, water; L, label; MRI, magnetic resonance imaging.

• 

  Figure 5. Comparison between raw MR images and pre-processed images of the (A) PDFF and (B) T2* modalities. As shown, noise in the background region is dramatically reduced. PDFF, proton density fat fraction; MR, magnetic resonance.

• 

  Figure 6. Comparison between (A) ground truth, predictions of (D-F) 2D U-Nets, (C) 3D U-Net, and (B) the BAT-Net. The yellow arrows mark the main difference compared to the proposed BAT-Net. BAT, brown adipose tissue; 3D, three-dimensional; 2D, two-dimensional.

• 

  Figure 7. Bland-Altman plots for comparison between annotations of two raters after label processing on (A) volume and (B) mean PDFF of the segmentation. Both volume and mean PDFF are in good agreement between different raters except for one striking outlier in the volume comparison. SD, standard deviation; PDFF, proton density fat fraction.

• 

  Figure 8. Bland-Altman plots for comparison between ground truth and predictions of the proposed BAT-Net on (A) volume and (B) mean PDFF of the segmentation. The volume segmented by the network is smaller on average compared to the ground truth labels while mean PDFF is increased. SD, standard deviation; PDFF, proton density fat fraction; BAT, brown adipose tissue.