DiffCAS

A myocardial perfusion SPECT study in a patient with a large chest or a high diaphragm frequently shows a thinned inferior wall that looks like ischaemia and is not: the gamma photons were absorbed by soft tissue on the way out. Hybrid SPECT/CT scanners resolve the ambiguity by measuring that absorption — a low-dose CT supplies voxel-wise linear attenuation coefficients the reconstruction divides back out. That hardware is expensive, adds dose, and is absent from most cameras in service, so the correction that would settle the reading is missing where cost pressure is greatest.

DiffCAS generates attenuation-corrected (AC) cardiac SPECT images directly from non-attenuation-corrected (NAC) reconstructions, with no CT at the point of use. It is built on a Brownian Bridge diffusion process, which models NAC→AC as a stochastic bridge between two image distributions rather than as denoising from pure Gaussian noise — the uncorrected study is one endpoint of the trajectory, not a conditioning vector bolted onto an unconditional generator. The physics is explicit: a reconstruction module predicts voxel-wise attenuation coefficients and path lengths, combines them through the Beer–Lambert law into a correction factor, and applies that factor at each diffusion step, pulling intermediate states toward physically consistent images rather than merely plausible ones.

It comes from the Institute for AI Innovation and Societal Impact (AI4LIFE) at Hanoi University of Science and Technology, with nuclear-medicine co-authors at 108 Military Central Hospital and a collaborator at Nagoya University, and succeeds the same group's PADM, which introduced both the teacher–student physics distillation scheme and the CardiAC dataset.

#Key Features

• CT-free deployment: A CT-conditioned teacher learns attenuation priors from paired NAC, CT and AC data; the student distilled from it runs on NAC alone — no CT hardware, no extra dose, no misregistration to quality-control.
• Beer–Lambert supervision inside the sampling loop: The predicted correction factor is applied at every denoising step, constraining the trajectory rather than scoring the output afterwards.
• Multi-slice contextual learning: Each sample spans the target slice together with eight neighbouring slices on either side, so cross-slice anatomy informs the correction and the volume stays spatially coherent.
• 3D CT Vision Transformer: A volumetric transformer models long-range structure in the CT attenuation map to supply the teacher's priors, which a slice-wise encoder cannot do for absorption paths through the thorax.
• Feature-level distillation: The student trains against an L1 reconstruction loss plus cosine-similarity alignment with the teacher's feature maps at three UNet stages, replicating internal representations rather than final pixels alone. A deformable registration network is co-trained with it to absorb NAC–AC slice misalignment.

#Technical Details

Training runs in two stages — CT-conditioned teacher first, then the CT-free student initialised against it — each for up to 200 epochs or 400,000 steps with Adam at a 1e-4 learning rate, using exponential moving average weights at 0.995 decay for validation and inference. The diffusion process uses 500 timesteps in training and 100 sampling steps at inference.

Both stages use CardiAC, a clinical corpus of 424 patient studies introduced by the PADM paper. Each study carries six paired NAC and AC reconstructions across rest and stress in the three standard nuclear-cardiology orientations — vertical long axis, horizontal long axis and short axis — plus two CT-derived attenuation maps. Acquisitions are technetium-99m MPI on GE SPECT systems at a 128×128 matrix with 25–49 slices per volume. Evaluation pairs global pixel-level image quality (SSIM, PSNR, RMSE and LPIPS) with myocardium-specific clinical metrics, against image-generative baselines on the same cohort.

#Applications

The immediate use is attenuation correction on SPECT-only cameras, which are the majority of installed myocardial perfusion scanners: a laboratory without CT can produce corrected short-axis and long-axis series as a preprocessing step, giving readers views they would otherwise infer from prone imaging or gated wall-motion data. Because the correction is generated rather than measured, it also sidesteps the SPECT–CT misregistration artefacts that need manual quality control on hybrid systems, and the corrected series feed the ordinary summed stress and rest scores.

#Impact

CT-free attenuation correction has been pursued with conditional GANs and U-Net regressors for several years; DiffCAS's contribution is to keep a diffusion model's generative flexibility while binding it to an explicit absorption model at each step, answering the standard objection that a network can hallucinate a clean myocardium with no physical warrant. Constraining the sampler with Beer–Lambert factors, then distilling those priors into a network that never sees a CT, is a pattern that transfers to other emission tomography settings where an anatomical prior exists in training but not in the clinic. The limits are real: evaluation is confined to the single CardiAC cohort on one vendor's systems, so cross-scanner behaviour is untested; no diagnostic-accuracy claim is made against invasive angiography; and the reference implementation ships without trained checkpoints, with CardiAC not yet downloadable, so reproduction needs comparable paired NAC/AC/CT data.