Fig 1.
Flowchart illustrating the major steps employed in segmentation and quantification of subcutaneous and visceral fat in MR images.
MR images were first converted to a format compatible with the ANTs software packages (NifTI; .nii.gz), and then processed for denoising and bias correction. 37 representative original MR images and their corresponding segmentation images from B6 mice and congenic mice were used for establishing the core training dataset. A novel data augmentation strategy was used to increase the number of training images from 37 to 60,552, which were used to train a U-net-based architecture. A testing dataset consisting of 60 MR images from 3 B6 mice was run through the model to test for accuracy of the method.
Fig 2.
Example MR images showing data augmentation strategy.
We introduced a novel data augmentation strategy for training through ANTs-based template construction. A template was created from the training data segmentation images where the red area includes visceral fat and the foreground designates regions of subcutaneous fat. 37 images were used to create such a template that permitting 372 = 1369 possible deformable shapes which are further augmented by random horizontal flipping and randomized rotation.
Fig 3.
Flowchart showing major steps in automated quantification of visceral and subcutaneous fat.
Images were created using the Fiji package of ImageJ. All MR images were set to the dimensions of 35x35 mm, a pixel width and height of 0.182291, and a voxel depth of 2. 1) Make the visceral segmentation image binary (outside = black, visceral component = white), and use the “Create Selection” tool to generate a selection perfectly outlining the visceral component. 2) Use the ‘Restore Selection” tool to place this selection on the original MR image, giving an exact outline of the visceral component. 3) Copy and paste this selected area to a new image, making sure it is the same dimensions as the original MR image to ensure accurate quantification (all images 35 x 35 mm; Pixel width/height = 0.182291; Voxel depth = 2.0). 4) Threshold the image to make a binary image in which the fat is one color and everything else is another. 5) Quantify this new image by using the “Analyze Particles” tool. 6) On the original MR image with the visceral fat component selected (from 2), use the “Make Inverse” tool to obtain the subcutaneous fat component. 7–9) Repeat the same quantification process as seen in 3–5.
Fig 4.
Body weight (g) of male congenic and B6-Apoe-/- mice fed a Western diet.
Results are means ± SE of 7 B6 and 17 congenic mice after 14 weeks of Western diet. * P < 0.05 vs. B6 mice.
Fig 5.
Representative MR images of congenic and B6-Apoe-/- mice fed a Western diet.
Axial MR slices at the levels denoted by the red (A) and green lines (B) across the coronal slice (C). D, gross examination of abdominal fat. Tope roll: B6-Apoe-/- mice; bottom: congenic mice.
Fig 6.
Fat and non-fat volumes of congenic and B6-Apoe-/- mice measured manually using axial MR slices.
Results are means ± SE of 4 mice per group. * P = 0.012.
Fig 7.
The accuracy of deep learning in deriving the area containing visceral fat at three different levels: pelvic, kidney and liver on MR images.
Prediction of the segmentation in the red area by deep learning is highly consistent with the input data. The red area is where visceral fat is included. Auto: prediction made by deep learning; manual: the red line is drawn with ImageJ and the area within the red line is used as input for visceral fat.
Fig 8.
The accuracy of deep learning in deriving the area containing subcutaneous fat at three different levels: Kidney and liver on MR images.
The red area is where subcutaneous fat is located. Auto: prediction made by deep learning; manual: the red line is drawn with ImageJ and the area outside the red line is used as input for subcutaneous fat.
Fig 9.
Comparison between the automated and manual measurements in quantification of visceral fat.
A, The volumes of visceral fat on 20 sequential axial slices from the tail root (slice 1) to the diaphragm (slice 20) of B6-Apoe-/- mice measured by manual (black) and deep learning (water filtered = solid; unfiltered = dashed). B, the total fat volume of B6-Apoe-/- mice measured by manual (black) and deep learning (grey; water filtered = solid, unfiltered = dashed). C, Correlation analysis of visceral fat volumes on water-filtered MR slices measured with the two methods. Each dot represents on axial MR slice. R2 and P values are shown in the figure. D, Correlation analysis of visceral fat volumes on unfiltered MR slices measured with the two methods.
Fig 10.
Comparison between the automated and manual measurements in quantification of subcutaneous fat.
A, The volumes of subcutaneous fat on 20 sequential axial slices from the tail root (slice 1) to the diaphragm (slice 20) of B6-Apoe-/- mice measured by a manual (black) or automated method (water filtered = solid; unfiltered = dashed). B, the total subcutaneous fat volume of B6-Apoe-/- mice measured by a manual (black) or automatic method (water filtered = solid; unfiltered = dashed). C, Correlation analysis of subcutaneous fat volumes on water-filtered MR slices measured with the two methods. Each dot represents a slice. R2 and P values are shown in the figure. D, Correlation analysis of subcutaneous fat volumes on unfiltered MR slices measured with the two methods.
Fig 11.
Comparison between congenic and B6-Apoe-/- mice in visceral (A) and subcutaneous fat volumes (B) measured by the automated method on water filtered MR slices. MR slices span from the pelvic cavity (slice 1) to the top of the liver (slice 20). Results are means ± SE of 4 mice per group on each slice. C, The total volume of visceral and subcutaneous fat in the abdominal region of congenic and B6-Apoe-/- mice. * P < 0.05.