search

dek924 / PerX2CT

Perspective Projection-Based 3D CT Reconstruction from Biplanar X-rays (ICASSP 2023, Best Student Paper Awards)
42 stars 9 forks source link

input at inference #4

Closed hyebinyoo closed 7 months ago

hyebinyoo commented 7 months ago

In this paper, you say "our model reconstructs the 3D CT volume using only two X-rays (PA and lateral view images) without additional information." But PerX2CT uses additional information 3d positional information that I think voxel position.

That x_i is the voxel position, that is, the 3D coordinates in the target CT. Don't you need that information when inferring?

hyebinyoo commented 7 months ago

And the paper says that "X-rays are generated by perspective projection". Do you have any related data? It is generally believed that X-rays are generated through orthogonal projection. And I can't find any data that says it's generated through perspective projection.

bttkm82 commented 7 months ago

x_i represents the voxel coordinates in a 3D CT volume, normalized relative to the volume's resolution, which is commonly correspond to a fixed size in previous studies (e.g., 128x128x128). Thus, it is not additional information like additional X-ray or semantic map. x_i is conceptually similar to positional encoding in Transformer models.

We recommend you to refer to the following papers about projection. SHEN, Liyue; ZHAO, Wei; XING, Lei. Patient-specific reconstruction of volumetric computed tomography images from a single projection view via deep learning. Nature biomedical engineering, 2019, 3.11: 880-888. Shen, L., Yu, L., Zhao, W., Pauly, J., & Xing, L. (2022). Novel-view X-ray projection synthesis through geometry-integrated deep learning. Medical image analysis, 77, 102372.