Predicting one-year left ventricular mass index regression following transcatheter aortic valve replacement in patients with severe aortic stenosis: A new era is coming

Figures

• 

  Figure 1. Schematic of the statistical shape analysis pipeline used for the outcome prediction.

• 

  Figure 2. Schematic representation of the U-Net neural network, semi-automatic segmentation of the LV using 3D-Slicer (blue), and a sample AI-predicted reconstructed LV geometry (red).

• 

  Figure 3. CNN architecture used for semantic image segmentation. The input and output images are shown as orange boxes. The inputs and outputs of the CNN layers are shown as green and blue boxes, respectively. The yellow boxes represent the copied tensors, and the blue arrow indicates the flow of the CNN model from raw input image to the binary segmented output image.

• 

  Figure 4. Some of 66 patient-specific left ventricle shapes after processing and preliminary operations. The template shape is indicated in the bottom right of the figure.

• 

  Figure 5. Percentage of variance explained in the outcome as a function of the number of shape components.

• 

  Figure 6. The first three shape modes which together convey 74% of the total covariance between the input and output. Along each row from left to right the shapes are ordered by increasing shape score as −2σi, −σi, 0, σi and 2σi, respectively. The template shape (s=0 along each mode) is indicated by the dashed box.

• 

  Figure 7. Predicted outcomes (vertical axis) vs clinically reported outcomes (horizontal axis) for train data (blue circles) and test data (red crosses). LVMI: Left ventricle mass index.

• 

  Figure 8. Applying linear regression surface to estimate post-LVMI based on clinically measured LVEF and pre-LVMI.

• 

  Figure 9. Critical points of heart contributing to the cardiac remodeling captured by the variable importance in projection (VIP) score which are all located on the internal wall of the left ventricle.