Generalizable Framework for Atrial Volume Estimation for Cardiac CT Images Using Deep Learning With Quality Control Assessment

Figures

• 

  Figure 1. The overview of the proposed framework for estimating the volume of LA.

• 

  Figure 2. The three steps of the Modified Reverse Classification Accuracy (mRCA) method.

• 

  Figure 3. Some examples of the prediction of the classification models. The top, middle, and bottom rows correspond to the predictions of the VGG16, VGG19, and ResNet50 models, respectively. The red arrows point to the LA. The middle column shows images with LA that the classification models got wrong.

• 

  Figure 4. The histogram (A) and the boxplot (B) of the dice score of the evaluation dataset. The 25, 50, and 75% percentiles are 0.88, 0.93, and 0.95, respectively.

• 

  Figure 5. Some examples of the variable prediction of the segmentation model (the red and blue contours represent the label and prediction and the corresponding dice scores are shown on the top of each image).

• 

  Figure 6. Examples of image quality assessment comparing RCA and mRCA methods. The “Good” (η ∈ [0 0.2]), “Average” (η ∈ [0.2 0.5]), and “Poor” (η ∈ [0.5 1.0]) refer to the quality of the predictions of the RCA method when compared to the actual DSC scores in terms of the relative error (not shown on the figure), where relative error η = |(DSCactual − DSCpredicted)/DSCactual|).

• 

  Figure 7. The regression (A,B) and the Bland-Altman (C,D) plots of the label volumes vs. the predicted volumes using the proposed framework. The data points are shown in blue and green in panel (A); the green data points (9 out of 337) in the figure have been flagged by the QC condition as poor LAV estimates therefore eliminated in panel (B). The red data point in panel (B) indicates one poor volume estimate (out of a total of 10) was missed by the imposed condition. In panels (C,D), the upper and lower dashed horizontal lines represent the confidence interval at 95%.

• 

  Figure 8. The top row shows (A) the regression plot and (B) the kernel density estimate plot along the two variables (label volume and predicted volume) marginal distributions. It also shows (C) the regression plot along the two variables' boxplots and (D) the kernel density estimate and histogram plots of the variables with the dashed vertical lines representing the arithmetic mean of the distributions. The symbols ρ and p represent the p-value and Pearson correlation coefficient, respectively. The plots (A–D) correspond to the label volumes vs. the predicted volumes without quality control (QC) and the plots (E–H) correspond to these variables using the proposed QC framework.