CT-free attenuation correction for cardiac SPECT perfusion scans, using a diffusion model whose every step applies a learned Beer–Lambert correction.
No providers recorded yet. Browse all providers
In a myocardial perfusion SPECT study, photons leaving the heart pass through different amounts of breast, diaphragm and bowel on their way to the detector, and the tissue that absorbs them leaves a dark patch in the image that a reader can mistake for a perfusion defect. Hybrid SPECT/CT scanners fix this by measuring the absorption directly: a low-dose CT yields voxel-wise attenuation coefficients that the reconstruction divides back out. Most installed cameras have no CT, and a network trained to translate uncorrected images into corrected ones has no physical reason to produce the right answer rather than a plausible one.
PADM (Physics-aware Attenuation Correction Diffusion Model) keeps the physics inside the generator. At every step of a Brownian-bridge diffusion process that runs from the non-attenuation-corrected (NAC) slice to its attenuation-corrected (AC) counterpart, a U-Net predicts an attenuation map and a set of per-angle photon path lengths, and those quantities are combined through the standard SPECT attenuation-correction factor — the Beer–Lambert exponential averaged over projection angles — and applied to the NAC image. The output is constrained to be the uncorrected slice multiplied by a physically parameterised correction, not an arbitrary image. CT enters only during training: a teacher network sees the CT-derived attenuation map, and a student with the same architecture learns to reproduce the teacher's behaviour from NAC input alone, so deployment needs no CT.
PADM was developed at the AI4LIFE institute of Hanoi University of Science and Technology, with nuclear-medicine co-authors at 108 Military Central Hospital and collaborators at Nagoya University and Japan's National Institute of Advanced Industrial Science and Technology. It was posted to arXiv in November 2025 and published at WACV 2026. The same group's DiffCAS extends it with multi-slice context and a volumetric CT teacher.
PADM operates on 2D slices in image space, without a VQGAN latent encoder: each slice is cropped to a 50×50 region around the heart and resized to 256×256. Training uses 500 diffusion timesteps, Adam at a 1e-4 learning rate and batch size 8 on a single 48 GB NVIDIA RTX A6000.
CardiAC contains 424 technetium-99m studies acquired in 2022–2023 on GE SPECT systems (323 male, 101 female; mean age 65.4), totalling 73,680 slices at a 128×128 matrix with 25–49 slices per volume and 128×128×128 attenuation maps. It was split by patient into 254 training, 84 validation and 86 test studies. On the test set the CT-free student reached an RMSE of 0.0218, SSIM of 0.9795 and PSNR of 34.75 dB, against 0.0217, 0.9796 and 34.98 dB for the CT-conditioned teacher. It had the lowest RMSE and highest PSNR among six baselines — Pix2Pix, RegGAN, ResViT, UNIT, BBDM and Palette — improving on the next-best method by 13.8% and 3.8% respectively, while UNIT, ResViT and RegGAN scored marginally higher on SSIM. The student–teacher gap widens as the number of projection angles in the correction factor rises from 16 to 64.
PADM targets nuclear-cardiology departments running SPECT-only cameras, where an attenuation-corrected series would otherwise require a hybrid scanner, extra radiation dose, or a repeat prone acquisition. Corrected short-axis and long-axis slices give readers a second view when judging whether an inferior or anterior wall defect is attenuation or ischaemia. CardiAC also serves as a benchmark for CT-free attenuation correction.
Earlier CT-free attenuation correction treated the problem as image translation with GANs or U-Nets, and diffusion models inherited the same weakness: nothing ties the output to how photons are actually absorbed. PADM's contribution is to parameterise the diffusion network's output through the attenuation-correction equation itself and to use distillation to remove the CT requirement at deployment. That design carried forward directly into DiffCAS. The evaluation is limited: all results come from a single cohort acquired on one vendor's systems, metrics are pixel-level image similarity rather than diagnostic accuracy, and the authors defer clinical reader studies to future work. Neither version of the paper links a code repository, trained checkpoints or a download location for CardiAC.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.