Efficient Pix2Vox++ for 3D Cardiac Reconstruction from 2D Echo Views

Figures

• 

  Fig. 1. Proposed pipeline for extracting 2D standard US views from a mesh and performing a reconstruction. A) Initial heart model with tissue labels, B) Mesh slicing with a plane and extracted segmentation-like images and C) Input images being trained/inferenced in the PiVox networks to produce 3D reconstructions.

• 

  Fig. 2. Example case of data generated at each stage in the process of creating a synthetic ultrasound image from a segmentation mask. From left to right, a) Initial segmentation image b) generated pseudo image c) generated synthetic US d) overlay of input mask on generated image.

• 

  Fig. 3. Comparison of the 4 networks and their theoretical total Multiply/Add operations for each number of input views at a 643 resolution.

• 

  Fig. 4. A) 9 binary views; B) Binary Apical 2 and 4 chamber views; C) CycleGAN inferenced Apical 2 and 4 chamber views; D) Example case from the CAMUS dataset. The same case is shown in A, B and C. The E-PiVox output column shows the ground truth (White) and the network predictions (Orange). The 3D Heart Anatomy column shows the absolute error in relation to the ground truth mesh (Black wireframe).