Benchmarks of Dr. TVAM

Here we show some simple benchmarks of Dr.TVAM.

Comparison of filtered backprojection vs Dr.TVAM

Reconstruction Results Comparison

All physical parameters are specified in this config file. Most notable, the attenuation coefficient is $\mu = 0.1 \frac{1}{\mathrm{mm}}$. The Radon transform assumes a $\mu \ll 1$, so in this case this requirement is violated (as in most realistic printing scenarios). Note, Dr.TVAM allows to control the sparsity of the patterns (over weight_sparsity in the config file). So even higher energy efficiencies can be achieved, if desired.

a), b), c), d) show different slices along the vertical axis of the final intensity and thresholded intensity.

Method	Dr. TVAM	Positive Filtered Backprojection
Intensity
Best Thresholded Intensity
Pattern for one angle
Histogram
IoU (1 is perfect)	0.9996	0.9325
Energy efficieny	0.0314	0.0037

Speed Benchmarks

Hardware Specifications

• RTX3060 12GB (2020): Entry-level consumer hardware GPU, priced around ~300USD.
• L40S 48GB (2022): Professional graphics card used primarily in server clusters, priced around 5000USD.

Test Subjects

• Dr. TVAM: Tested on RTX3060 12GB and L40S 48GB.

• Benchy boat as target:

  • Size: 10mm
  • Pixel size: 25µm
  • Resolution on DMD: 400x400 pixels
  • Angles: 400
  • 40 iterations with gradient-based L-BFGS optimizer

• The Julia based (attenuated) Radon transform is using the same parameters and optimizer.

Computation Performance

Configuration	Rays per Pixel	Time (RTX3060)	Time (L40S)
Single backprojection (adjoint Radon) in Julia	1	0h 0m 1s
Julia (attenutaed) Radon based	1	0h 3m 17s
Dr. TVAM index-matched	1	0h 2m 2s	0h 0m 20s
Dr. TVAM cylindrical	1	0h 2m 10s	0h 0m 23s
Dr. TVAM square	1	0h 2m 15s	0h 0m 20s
Dr. TVAM cylindrical scattering	16	1h 40m 0s	0h 14m 10s
Dr. TVAM square scattering	16	1h 40m 0s	0h 14m 20s
Dr. TVAM square scattering (surface-aware loss, disable black pixels)	16	0h 25m 0s	0h 3m 45s