3 ms·

CT scan time is ~0.2-0.5 seconds for a 160mm slice, so a whole body can probably be scanned fairly quickly. However, CT scanning is only done for the body parts

by car 12y ago

CT scan time is ~0.2-0.5 seconds for a 160mm slice, so a whole body can probably be scanned fairly quickly. However, CT scanning is only done for the body parts of interest, to keep radiation exposure low. Most fascinating here is the ability to capture a beating heart!

ahelwer 12y ago
It's also computationally intensive to reconstruct all the x-ray images into a 3D model. I worked for a company that wrote software to speed this up with GPGPUs. Standard software packages took on the order of hours for one scan, while parallelism brought it down to 10s of minutes - this was just to transform the x-ray data into voxels. Not sure how long the post-processing to identify & highlight organs would take. It doesn't matter so much for humans, though. The main market for the above-mentioned software product was research labs imaging hundreds of lab rats.
- marvin 12y ago
  The pretty pictures in the article can be rendered in real time using GPUs. Unless the pictures here use techniques that I haven't heard about, it is done by treating the voxels as a semi-transparent cloud of different-colored materials and using a fragment/pixel shader to simulate rays of light projected through this volume. It's pretty straightforward; the volume data is stored in a 3D texture and each ray of light is simulated using one "pixel" of the fragment shader. This is enhanced by assigning more weight to the voxels that are in a boundary region between tissues of different density (you see this effect in the pictures of lungs, kidneys etc; the organ is transparent but the edges are visible) and applying a shading model to these boundary regions to improve perception of their spatial orientation. Different organs can be pre-selected by marking their constituent voxels, and different models/colors/shading applied to each organ. This process can also sometimes be automated, if you have the luxury of a table that correlates voxel intensity (x-ray absorption value) with tissue type (bone, muscle, air, etc).
  - ajuc 12y ago
    I thought even to get voxels you first have to "solve sudoku"? You have sum of all the voxels "opacity" along rays from many directions and you want to know each voxel opacity.
    - marvin 12y ago
      Yes, my comment only treats the process from the point where you have a complete dataset of voxels to the point where you have a finished rendering. ahelwer didn't treat this part of the process in his comment.
    - ahelwer 12y ago
      It's much, much simpler than that. You just project a ray from the X-ray point source through the center of a voxel onto the detector, interpolate the X-ray energy detected, and add that energy to the sum in that voxel. Repeat for each voxel and each x-Ray taken (CT scans take a bunch of x-Rays going around in a circle), remove some noise, and boom you've voxelled the data. Fundamentally you're just summing the energy that passes through each point in space. It's trivially parallelizeable.

by car 12y ago

ahelwer 12y ago

It's also computationally intensive to reconstruct all the x-ray images into a 3D model. I worked for a company that wrote software to speed this up with GPGPUs. Standard software packages took on the order of hours for one scan, while parallelism brought it down to 10s of minutes - this was just to transform the x-ray data into voxels. Not sure how long the post-processing to identify & highlight organs would take. It doesn't matter so much for humans, though. The main market for the above-mentioned software product was research labs imaging hundreds of lab rats.

marvin 12y ago
The pretty pictures in the article can be rendered in real time using GPUs. Unless the pictures here use techniques that I haven't heard about, it is done by treating the voxels as a semi-transparent cloud of different-colored materials and using a fragment/pixel shader to simulate rays of light projected through this volume. It's pretty straightforward; the volume data is stored in a 3D texture and each ray of light is simulated using one "pixel" of the fragment shader. This is enhanced by assigning more weight to the voxels that are in a boundary region between tissues of different density (you see this effect in the pictures of lungs, kidneys etc; the organ is transparent but the edges are visible) and applying a shading model to these boundary regions to improve perception of their spatial orientation. Different organs can be pre-selected by marking their constituent voxels, and different models/colors/shading applied to each organ. This process can also sometimes be automated, if you have the luxury of a table that correlates voxel intensity (x-ray absorption value) with tissue type (bone, muscle, air, etc).
- ajuc 12y ago
  I thought even to get voxels you first have to "solve sudoku"? You have sum of all the voxels "opacity" along rays from many directions and you want to know each voxel opacity.
  - marvin 12y ago
    Yes, my comment only treats the process from the point where you have a complete dataset of voxels to the point where you have a finished rendering. ahelwer didn't treat this part of the process in his comment.
  - ahelwer 12y ago
    It's much, much simpler than that. You just project a ray from the X-ray point source through the center of a voxel onto the detector, interpolate the X-ray energy detected, and add that energy to the sum in that voxel. Repeat for each voxel and each x-Ray taken (CT scans take a bunch of x-Rays going around in a circle), remove some noise, and boom you've voxelled the data. Fundamentally you're just summing the energy that passes through each point in space. It's trivially parallelizeable.