FoundDiff: Unified LDCT Denoising

• The paper introduces FoundDiff, a two-stage denoising framework that couples dose-anatomy perception via DA-CLIP with adaptive restoration through DA-Diff.
• FoundDiff models dose as a continuous, ordinal condition and anatomy as a semantic variable, using distinct conditioning mechanisms to enhance denoising performance.
• The framework achieves superior generalization with improved PSNR/SSIM metrics on both seen and unseen dose levels compared to specialized, single-condition models.

FoundDiff is a two-stage, unified low-dose computed tomography (LDCT) denoising framework designed to work across multiple dose levels and multiple anatomical regions without retraining or manually supplying explicit condition labels at test time. It couples dose-anatomy perception through a dose- and anatomy-aware contrastive language image pre-training model (DA-CLIP) with adaptive denoising through a dose- and anatomy-aware diffusion model (DA-Diff), and is presented as a foundational diffusion model for unified and generalizable LDCT denoising across various dose levels and anatomical regions (Chen et al., 24 Aug 2025).

1. Clinical problem and scope

FoundDiff is motivated by the central tension of LDCT imaging: reduced radiation exposure is desirable under the ALARA principle, but lower radiation increases noise and artifacts and can compromise diagnostic image quality. The restoration objective is therefore to recover an NDCT-quality image from an LDCT input. The paper treats this not as a single-condition denoising problem, but as a setting in which noise severity varies continuously with dose and image statistics vary substantially across anatomical regions such as abdomen, chest, and head (Chen et al., 24 Aug 2025).

The paper argues that most prior DL-based LDCT denoisers are trained narrowly, usually on one dose level or one anatomy, and therefore learn dose- and anatomy-specific mappings. It identifies several concrete limitations: single-condition training, discrete explicit dose conditioning, neglect of anatomy semantics, lack of adaptive perception, and the scalability cost of specialized models. In this formulation, generalization failure is expected whenever scanning conditions shift across dose protocols or anatomical regions, because the model has not explicitly learned to perceive those variations (Chen et al., 24 Aug 2025).

A central conceptual point is the paper’s use of the term “foundational diffusion model.” Here the phrase denotes a unified, broadly applicable LDCT denoiser trained across multiple dose levels and multiple anatomical regions so that it can serve as a general-purpose denoiser with strong transfer, especially to unseen dose levels, without extra fine-tuning. It does not denote a web-scale multimodal foundation model; rather, it denotes a single denoising framework intended to subsume condition-specific LDCT models (Chen et al., 24 Aug 2025).

2. Two-stage architecture

FoundDiff is organized as a sequential two-stage strategy. Stage 1 is dose-anatomy perception through DA-CLIP. Stage 2 is adaptive denoising through DA-Diff. The interaction is direct: DA-CLIP takes a CT image and produces a dose embedding $e_\mathrm{d}$ and an anatomical embedding $e_\mathrm{a}$; DA-Diff then uses those learned embeddings to condition a diffusion denoiser that restores the LDCT image (Chen et al., 24 Aug 2025).

At inference time, the input is an LDCT image only. The image is first passed through DA-CLIP, which outputs dose-aware and anatomy-aware embeddings. These embeddings condition DA-Diff, which denoises the LDCT and produces the restored image. The paper explicitly emphasizes that this avoids manually supplying explicit condition labels at test time. At training time, the two stages are optimized sequentially: DA-CLIP is trained first on images with dose labels and anatomy labels, and DA-Diff is then trained on LDCT/NDCT pairs while using the learned DA-CLIP representations as conditioning (Chen et al., 24 Aug 2025).

This decomposition assigns distinct roles to the two stages. DA-CLIP performs perception of scanning-condition semantics, with dose treated as a continuous ordinal variable and anatomy as a region-level semantic variable. DA-Diff performs restoration in residual space, using those condition embeddings to modulate denoising behavior. The framework therefore combines explicit condition perception with conditional diffusion, rather than relying on a single monolithic denoiser to infer all condition structure implicitly (Chen et al., 24 Aug 2025).

3. DA-CLIP: dose and anatomy perception

DA-CLIP is built on a pretrained CLIP-based image quality assessment model, specifically CLIPIQA, and fine-tunes the image encoder while adding two task-specific heads: a two-layer MLP dose head producing $e_\mathrm{d}$, and a two-layer MLP anatomical head producing $e_\mathrm{a}$ (Chen et al., 24 Aug 2025).

For dose perception, the model also uses two text embeddings, $e_\text{clean}$ and $e_\text{noisy}$, obtained from the prompts “clean” and “noisy.” The predicted dose score is

$$\hat{y}_\mathrm{d}= \frac{\exp(e_\text{d}\cdot e_\text{clean})} {\exp(e_\text{d}\cdot e_\text{clean})+\exp(e_\text{d}\cdot e_\text{noisy})}.$$

This scalar is supervised by

$$\mathcal{L}_\text{dose}=\|\hat{y}-y_\mathrm{d}\|_2^2,$$

where $y_\mathrm{d}$ is the ground-truth dose label. The paper argues that MSE alone produces fragmented, dispersed representations and therefore introduces a ranking contrastive loss $\mathcal{L}_\mathrm{rank}$, constructed with $e_\mathrm{a}$0 as the $e_\mathrm{a}$1 distance between dose labels and a contrastive set $e_\mathrm{a}$2 defined by relative dose distance. The resulting dose objective is

$e_\mathrm{a}$3

Its stated purpose is to encode the continuous and ordinal nature of dose variation, which is especially important because the dose levels are concentrated in a narrow low-dose range from $e_\mathrm{a}$4 to $e_\mathrm{a}$5 of normal dose (Chen et al., 24 Aug 2025).

For anatomy perception, the model uses a supervised contrastive loss over the anatomical embedding: $e_\mathrm{a}$6 Here $e_\mathrm{a}$7 is the set of positives from the same anatomical region. The total DA-CLIP objective is

$e_\mathrm{a}$8

The anatomical classes used are abdomen, chest, and head (Chen et al., 24 Aug 2025).

The empirical evidence reported for DA-CLIP is strong. For dose prediction, CLIPIQA gives PLCC/SROCC of $e_\mathrm{a}$9, CLIPIQA+ gives $e_\mathrm{d}$0, DA-CLIP without $e_\mathrm{d}$1 gives $e_\mathrm{d}$2, DA-CLIP without $e_\mathrm{d}$3 gives $e_\mathrm{d}$4, and full DA-CLIP gives $e_\mathrm{d}$5. The paper also reports t-SNE behavior consistent with the intended geometry: dose embeddings are arranged continuously according to relative dose level, whereas anatomy embeddings form clear clusters for abdomen, chest, and head (Chen et al., 24 Aug 2025).

4. DA-Diff and DACB-conditioned residual diffusion

DA-Diff is built on RDDM and uses the residual diffusion branch. Instead of predicting a clean image directly, it operates in residual space, defining

$e_\mathrm{d}$6

where $e_\mathrm{d}$7 is the LDCT image and $e_\mathrm{d}$8 is the NDCT image. The forward residual diffusion process is

$e_\mathrm{d}$9

and the training objective is

$e_\mathrm{a}$0

The network itself is U-Net-like, with four levels and channel widths $e_\mathrm{a}$1. Each level contains an RLEB for local contextual information and a DACB for condition-aware modulation (Chen et al., 24 Aug 2025).

DACB is the distinctive conditioning mechanism. The paper assigns different roles to the two conditions. Dose is treated as a global condition, analogous to diffusion timestep, and is fused with timestep through adaLN-Zero. Anatomy is treated as spatial-semantic context and is integrated through a conditional state-space model derived from Mamba, denoted CSSM. Dose-timestep modulation is parameterized as

$e_\mathrm{a}$2

and the two DACB sub-blocks are written as

$e_\mathrm{a}$3

$e_\mathrm{a}$4

followed by

$e_\mathrm{a}$5

$e_\mathrm{a}$6

The paper states that the anatomical embedding is linearly projected and then added to the SSM output matrix $e_\mathrm{a}$7 inside the conditional 2D selective scan, while transposed attention is used to mitigate the channel-independence limitation of standard SSM processing (Chen et al., 24 Aug 2025).

The DA-Diff ablation starting from RDDM shows a monotonic improvement as the proposed conditioning components are added. Reported PSNR/SSIM values are $e_\mathrm{a}$8 for RDDM, $e_\mathrm{a}$9 after adding dose condition, $e_\text{clean}$0 after adding anatomical condition through CSSM, $e_\text{clean}$1 after adding transposed attention, and $e_\text{clean}$2 for the full FoundDiff configuration. This supports the paper’s claim that dose and anatomy should be fused differently and jointly (Chen et al., 24 Aug 2025).

5. Data regime, optimization, and empirical performance

The main training and evaluation corpus is a large simulated dataset derived from Mayo-2020. It uses 50 patients per anatomical region for abdomen, chest, and head, and extends the original low-dose availability by simulating additional dose levels with the ASTRA Toolbox. The resulting eight dose levels are $e_\text{clean}$3. The patient split is $e_\text{clean}$4 for training and $e_\text{clean}$5 for testing. Across all doses, the training set contains 85,480 slices and the testing set contains 21,424 slices; per dose, the totals are 10,685 training slices and 2,678 testing slices (Chen et al., 24 Aug 2025).

Dataset	Role	Key facts
Mayo-2020 derived set	Main training/testing	8 dose levels, 3 anatomical regions, 50 patients per region
Mayo-2016	Cross-dataset generalization	Abdomen only, 10 patients, 2 patients for testing, 1,136 images

DA-CLIP is trained on all data. DA-Diff is trained only on four seen dose levels, $e_\text{clean}$6, and is then tested on both those seen doses and the unseen doses $e_\text{clean}$7. This is the paper’s operational definition of unseen-dose generalization (Chen et al., 24 Aug 2025).

The optimization settings are explicit. DA-CLIP is trained for 100 epochs with batch size 192 using SGD, momentum $e_\text{clean}$8, weight decay $e_\text{clean}$9, and a cosine-annealed learning rate from $e_\text{noisy}$0 to $e_\text{noisy}$1. DA-Diff is trained for 400K iterations with batch size 2 using Adam with $e_\text{noisy}$2, $e_\text{noisy}$3, weight decay $e_\text{noisy}$4, and learning rate $e_\text{noisy}$5. DA-CLIP uses $e_\text{noisy}$6 patches; DA-Diff uses full $e_\text{noisy}$7 images. Window level is $e_\text{noisy}$8 HU, and augmentation consists of random horizontal flip and rotations $e_\text{noisy}$9. Diffusion uses $$\hat{y}_\mathrm{d}= \frac{\exp(e_\text{d}\cdot e_\text{clean})} {\exp(e_\text{d}\cdot e_\text{clean})+\exp(e_\text{d}\cdot e_\text{noisy})}.$$0 total steps and DDIM sampling with only 2 sampling steps (Chen et al., 24 Aug 2025).

On seen dose levels, the paper reports that FoundDiff consistently outperforms all competing methods on PSNR and SSIM. On unseen dose levels $$\hat{y}_\mathrm{d}= \frac{\exp(e_\text{d}\cdot e_\text{clean})} {\exp(e_\text{d}\cdot e_\text{clean})+\exp(e_\text{d}\cdot e_\text{noisy})}.$$1, it also outperforms competing methods, with especially clear gains at $$\hat{y}_\mathrm{d}= \frac{\exp(e_\text{d}\cdot e_\text{clean})} {\exp(e_\text{d}\cdot e_\text{clean})+\exp(e_\text{d}\cdot e_\text{noisy})}.$$2. The qualitative description is consistent across abdomen, chest, and head: FoundDiff better preserves vessels and fine structures, suppresses noise without artificial textures, avoids over-smoothing, and better matches NDCT reference (Chen et al., 24 Aug 2025).

The most concrete quantitative comparison is cross-dataset testing on Mayo-2016. FoundDiff attains $$\hat{y}_\mathrm{d}= \frac{\exp(e_\text{d}\cdot e_\text{clean})} {\exp(e_\text{d}\cdot e_\text{clean})+\exp(e_\text{d}\cdot e_\text{noisy})}.$$3 PSNR and $$\hat{y}_\mathrm{d}= \frac{\exp(e_\text{d}\cdot e_\text{clean})} {\exp(e_\text{d}\cdot e_\text{clean})+\exp(e_\text{d}\cdot e_\text{noisy})}.$$4 SSIM, compared with $$\hat{y}_\mathrm{d}= \frac{\exp(e_\text{d}\cdot e_\text{clean})} {\exp(e_\text{d}\cdot e_\text{clean})+\exp(e_\text{d}\cdot e_\text{noisy})}.$$5 and $$\hat{y}_\mathrm{d}= \frac{\exp(e_\text{d}\cdot e_\text{clean})} {\exp(e_\text{d}\cdot e_\text{clean})+\exp(e_\text{d}\cdot e_\text{noisy})}.$$6 for RDDM, $$\hat{y}_\mathrm{d}= \frac{\exp(e_\text{d}\cdot e_\text{clean})} {\exp(e_\text{d}\cdot e_\text{clean})+\exp(e_\text{d}\cdot e_\text{noisy})}.$$7 and $$\hat{y}_\mathrm{d}= \frac{\exp(e_\text{d}\cdot e_\text{clean})} {\exp(e_\text{d}\cdot e_\text{clean})+\exp(e_\text{d}\cdot e_\text{noisy})}.$$8 for Restormer, $$\hat{y}_\mathrm{d}= \frac{\exp(e_\text{d}\cdot e_\text{clean})} {\exp(e_\text{d}\cdot e_\text{clean})+\exp(e_\text{d}\cdot e_\text{noisy})}.$$9 and $$\mathcal{L}_\text{dose}=\|\hat{y}-y_\mathrm{d}\|_2^2,$$0 for CoreDiff, $$\mathcal{L}_\text{dose}=\|\hat{y}-y_\mathrm{d}\|_2^2,$$1 and $$\mathcal{L}_\text{dose}=\|\hat{y}-y_\mathrm{d}\|_2^2,$$2 for RED-CNN, and $$\mathcal{L}_\text{dose}=\|\hat{y}-y_\mathrm{d}\|_2^2,$$3 and $$\mathcal{L}_\text{dose}=\|\hat{y}-y_\mathrm{d}\|_2^2,$$4 for PDF (Chen et al., 24 Aug 2025).

A second notable result is comparison against specialized single-condition models at $$\mathcal{L}_\text{dose}=\|\hat{y}-y_\mathrm{d}\|_2^2,$$5 dose. FoundDiff still slightly exceeds those specialized variants: on abdomen it reaches $$\mathcal{L}_\text{dose}=\|\hat{y}-y_\mathrm{d}\|_2^2,$$6 versus RDDM $$\mathcal{L}_\text{dose}=\|\hat{y}-y_\mathrm{d}\|_2^2,$$7; on chest it reaches $$\mathcal{L}_\text{dose}=\|\hat{y}-y_\mathrm{d}\|_2^2,$$8 versus Restormer $$\mathcal{L}_\text{dose}=\|\hat{y}-y_\mathrm{d}\|_2^2,$$9; and on head it reaches $y_\mathrm{d}$0 versus RDDM $y_\mathrm{d}$1. This is important because it shows that unification across dose and anatomy does not require sacrificing specialized-condition performance (Chen et al., 24 Aug 2025).

6. Interpretation, limitations, and relation to other “FoundDiff” usages

The most important technical interpretation of FoundDiff is that it models dose as an ordinal continuous condition and anatomy as a semantic structural condition, then fuses them with different mechanisms: adaptive normalization for dose/timestep and conditional state-space modeling for anatomy. This is the core reason the paper gives for its unseen-dose behavior. The DA-CLIP ablations support the conditioning hypothesis directly: removing $y_\mathrm{d}$2 reduces dose prediction to PLCC/SROCC $y_\mathrm{d}$3, and removing $y_\mathrm{d}$4 yields $y_\mathrm{d}$5, whereas full DA-CLIP reaches $y_\mathrm{d}$6 (Chen et al., 24 Aug 2025).

The paper is also explicit about limits. It notes a simulation-to-real gap, because the simulated dataset still differs from heterogeneous real clinical LDCT data. It also notes computational cost: even with efficient SSMs and 2-step DDIM sampling, full-image diffusion denoising remains more expensive than traditional models. Additional constraints are visible from the setup itself: DA-CLIP requires dose labels and anatomy labels during training, and the anatomical coverage in the reported experiments is limited to abdomen, chest, and head (Chen et al., 24 Aug 2025).

A common source of confusion is the name. In the supplied literature, similarly named or conceptually adjacent systems address different tasks. Diff-XYZ studies code-diff understanding through apply, anti-apply, and diff generation over $y_\mathrm{d}$7 triples (Glukhov et al., 14 Oct 2025). DiffRegCD uses frozen denoising diffusion features for integrated registration and change detection, estimating both a dense displacement field and a change mask (Madani et al., 11 Nov 2025). DifFoundMAD applies foundation-model embeddings to differential morphing attack detection by comparing a suspected morph and a trusted live capture (Gonzalez-Soler et al., 20 Apr 2026). FoundDiff, by contrast, is specifically a unified LDCT denoising framework. A plausible implication is that the shared “FoundDiff” vocabulary in these works reflects a broader pattern—foundation representations or diffusion-based priors being coupled to differential or cross-condition reasoning—but the concrete object of inference in FoundDiff is image restoration rather than correspondence, biometric discrepancy, or code-edit reasoning (Chen et al., 24 Aug 2025).

In that sense, FoundDiff’s contribution is best understood narrowly and technically: it is a two-stage LDCT denoising system in which condition perception and conditional residual diffusion are explicitly separated, and in which generalization across eight dose levels and three anatomical regions is treated as the primary design objective rather than as a secondary transfer property (Chen et al., 24 Aug 2025).