Automated Quality-Controlled Cardiovascular Magnetic Resonance Pericardial Fat Quantification Using a Convolutional Neural Network in the UK Biobank

Figures

• 

  Figure 1. Two examples of PAT contoured in end-diastole on four-chamber bSSFP cine-CMR, performed using CVI42® software according to the established SOP. A single contour was drawn to select areas of high signal intensity adjacent to the epicardial surface of the left and right ventricles, resulting in output of an area measure (A). Areas of high signal intensity over the liver were not included in the PAT measure as this almost always represents adipose tissue below the diaphram (B). bSSFP, balanced steady state free precession; CMR, cardiovascular magnetic resonance; PAT, pericardial adipose tissue; SOP, standard operating procedure. Images reproduced with permission of UK Biobank.

• 

  Figure 2. Central illustration. Summary of model architecture used in the present study. The MultiRes blocks (A) form the encoder and decoder arms of the network. The number of filters used throughout the different components of the block is parameterized by U. The encoder and decoder arms are joined by Res paths (B), which are parameterized by F and L. They are formed of L repeating units, and their convolutional components each have F filters. The complete network is shown in (C). In (C), the colors indicate the placement of MultiRes blocks (A) and Res paths (B), while the hyperparameters used in each instance of the blocks are indicated as overlaid text. Because of the permanently active dropout components, each prediction the network makes is equivalent to a Monte Carlo (MC) sample. (D) shows three such samples drawn based on the same input image. Note the disagreement at the edges of the segmented regions, particularly clear as shown on the overlay (far right). Images reproduced with permission of UK Biobank.

• 

  Figure 3. Model performance. (A–D) Histograms of standard segmentation performance metrics on the test set (n = 87). (E–H) Bland–Altman plots of PAT area between manual measurement between measurements by different human observers, and a human observer (O1) and automated measurement. The x-axis denotes the average of two measurements and the y-axis denotes the difference between them. The dark line is the mean difference, and the dashed lines show ±1.96 standard deviations from the mean. (E–G) show the inter-observer variability evaluated by three observers (O1–O3) on a randomly selected subset of the manually contoured training set (n = 50 subjects). (H) shows the agreement between automated and manual measurements in the manually contoured test set (n = 87 subjects). (I) Example segmentations from the test set, with annotations showing Dice score and the predicted Dice score. Images reproduced with permission of UK Biobank.

• 

  Figure 4. Comparison of quantified PAT from CT and CMR. (A) The predicted Dice scores of the segmented data. (B) Correlation between PAT volume quantified via QFAT software and PAT area quantified using our method. Subjects with a predicted Dice <0.6 were excluded from Pearson analysis. (C) Some example CMR images, their automatically segmented PAT, and the predicted segmentation quality are also shown for reference.