Press “Process image” to see the result here
Residual = acquired minus model, added back as texture: 0% fully denoised, 50% for measurements, 20–30% for display. Both images share one window.
AcquiredProcessedClick either image to choose the profile row
DICOM in whole counts with the same total; no patient details.
How it works
- Variance stabilisation. The Anscombe transform turns Poisson counts into data with near-constant noise (σ ≈ 1).
- BM3D denoising. Similar 8 × 8 patches are grouped and filtered together, averaged over four shifted runs so block edges never line up. Below about 2 counts per pixel, a count-adaptive smoother that uses only background pixels takes over.
- Exact unbiased inverse. The Mäkitalo–Foi inverse returns the image to counts without the low-count bias of the simple algebraic inverse.
- Richardson–Lucy deconvolution. A 4.5 mm Gaussian kernel, 40 iterations (20 for pixels under 2 mm), with a floor at the local background so edges cannot dip below it. This removes the dark halo around hot structures.
- Residual restoration. A chosen fraction of the residual (acquired minus model) is added back, rising to 100% in empty background. Total counts are rescaled to match the acquired image.
Validated settings
| Parameter | Value |
|---|---|
| Camera / collimator | GE Discovery NM630, LEHR |
| Radionuclide | Tc-99m (also tested on I-131) |
| Denoising grid | ≈ 2.2 mm (finer images are binned) |
| BM3D noise σ | 1.0 (Anscombe domain) |
| RL kernel | 4.5 mm FWHM Gaussian |
| RL iterations | 40 (20 if pixel < 2 mm) |
| Low-count blend | 2 counts / pixel |
| Fraction of residual | 50% quantitative, 20–30% display |
Use with care
- For research and education. Not a medical device, and not for diagnosis.
- Load raw count images (NM DICOM), not screenshots or already-filtered images. The method relies on Poisson statistics.
- Settings were optimised on Tc-99m with LEHR. Other radionuclides and collimators work, but the kernel and fraction of residual should be re-checked for each.
- Post-processing cannot recover detail the camera never recorded. Very small, low-count lesions can lose contrast-to-noise; always compare with the acquired image.
- This browser version of BM3D (DCT and Haar transforms) differs slightly from the Python reference. Outputs agree to about 6% RMS, which is less than half of a 50% residual.
Cite this work
Mathiyazhagan P, Jeba S A. From Stars to Scintigrams: an astronomy-inspired resolution recovery and denoising algorithm for planar gamma camera imaging. Poster, FASMI Biannual Conference 2026, Hyderabad. Department of Nuclear Medicine, JIPMER, Puducherry.
Phantom samples: Jaszczak phantom with hot inserts, four-quadrant bar phantom and Tc-99m line source, acquired on a GE NM630 at JIPMER. Only the count values are included; no headers.