FoundDiff

Noise in a CT scan is not one thing. A half-dose abdominal acquisition and a tenth-dose head scan differ in both photon statistics and the anatomy underneath, and a denoising network trained on the first degrades visibly on the second. The field's working answer has been a grid of specialists — one model per dose level, per body region — which multiplies training cost, and still has nothing to offer the scanning protocol nobody anticipated. Nominal dose settings are not even consistent across vendors, so the label such a model conditions on need not match the noise it receives.

FoundDiff replaces that grid with one checkpoint that infers a scan's dose and anatomy from the image itself and adapts its denoising accordingly. It was developed at the Institute of Science and Technology for Brain-inspired Intelligence at Fudan University, with collaborators at Sichuan University and Shanghai Jiao Tong University, and published in IEEE Transactions on Medical Imaging.

The design runs in two stages. A dose- and anatomy-aware CLIP model (DA-CLIP) produces a continuous dose embedding and a clustered anatomical embedding; a residual diffusion model (DA-Diff) consumes both as conditioning. Where CT Foundation compresses a reconstructed volume for downstream interpretation, FoundDiff works on the image itself, filling the restoration role that S³-Diff occupies for histology.

#Key Features

• Dose read as a continuous quantity: DA-CLIP is fine-tuned from a CLIP-based image quality assessment backbone with a dose-ranking contrastive loss, so embeddings preserve the ordinal structure of dose rather than snapping to discrete class labels. Dose prediction reaches a Pearson correlation of 0.9907 and a Spearman correlation of 0.9831.
• Anatomy discrimination alongside it: a supervised contrastive loss over abdomen, chest and head clusters the anatomical embedding, so the conditioning signal carries where the noise sits as well as how much of it there is.
• Mamba-based conditional block: the dose and anatomy conditional block (DACB) fuses the dose embedding with the diffusion timestep through adaptive layer normalization for global modulation, and injects anatomy through a conditional state-space model whose 2D selective scan captures long-range spatial context at linear complexity.
• No explicit condition at inference: because perception is learned from the image, FoundDiff needs no dose input, no protocol string and no fine-tuning when it meets an unfamiliar scanner or region.
• Transfer to unseen dose levels: 1/7 and 1/15 of normal dose are withheld from every training stage, and FoundDiff leads on both across all three anatomical regions.

#Technical Details

Training uses a simulated corpus built from the 2020 AAPM-Mayo projection dataset: 50 patients across abdomen, chest and head on GE Discovery and Siemens SOMATOM scanners, with the ASTRA Toolbox used to synthesize ten dose levels from 1/2 down to 1/20 under a Poisson-plus-Gaussian measurement model. That yields 13,363 slices (5,339 abdomen, 6,616 chest, 1,408 head) and 133,630 slices across all dose levels. DA-CLIP sees eight dose levels; DA-Diff is trained on only four. The denoising network follows the residual diffusion (RDDM) formulation in a four-level U-Net with 64/128/256/512 channels, each level pairing a residual local-enhance block with a DACB, at 21.6M parameters — comparable to the RDDM baseline it builds on — and sampling collapses 1,000 diffusion steps to two, giving 0.327 seconds per 512×512 image.

Cross-dataset evaluation tests the generalization claim. On the multi-centre CQ500 head dataset (six Indian radiology centres, GE and Philips scanners), FoundDiff reaches 56.88 dB PSNR at 1/4 dose and still leads at the strictly unseen 1/7 and 1/15 levels. On the piglet dataset — real reduced-tube-current scans, not simulation — it is best on every metric at all four dose levels, including 38.59 dB at 1/20. Against specialists trained for a single condition, the unified model still wins: 45.94 dB versus 45.61 for a condition-specific RDDM on 1/10-dose abdomen.

#Applications

The target is clinical CT reconstruction under the ALARA principle, where dose is pushed as low as diagnostic quality permits and protocols differ by patient age, target anatomy and vendor. One checkpoint that accepts any reconstructed image and infers its own conditioning suits PACS integration and multi-site deployment, where maintaining a separate model per protocol is impractical, and needing no projection data means it applies to archived reconstructions as well as live acquisitions.

#Impact

FoundDiff is a concrete demonstration that the "one model per setting" convention in low-dose CT restoration is avoidable, and that perceptual conditioning learned from the image can substitute for explicit dose metadata that is often unreliable. The authors are candid about the limits: the main training corpus is simulated, so a simulation-to-real gap persists despite the piglet and CQ500 results; diffusion sampling remains more expensive than a feed-forward CNN; and restorations at ultra-low dose can look mildly over-smoothed. Trained weights for both stages are released, alongside the model code and the DA-Diff training loop, though the stage-one DA-CLIP training procedure is absent from the repository, which carries no license file.