ALL-PET: Low-Shot PET Model in Projection Domain

• ALL-PET is a low-shot PET foundation model that operates on raw sinograms using latent diffusion to preserve acquisition geometry and address data scarcity.
• It employs Radon mask augmentation and dynamic multi-mask generation to yield over 200,000 diverse training samples from just 500 sinograms.
• The approach integrates parameter-free transparent medical attention that leverages PET-specific geometric constraints for improved interpretability and diagnostic guidance.

Searching arXiv for papers on ALL-PET and closely related PET foundation-model work. I’m looking up relevant arXiv records for ALL-PET and adjacent PET foundation-model papers. ALL-PET most directly denotes a low-resource, low-shot PET foundation model that operates in the projection domain, rather than on reconstructed PET images, and is built around a latent diffusion model applied to raw sinograms (Huang et al., 11 Sep 2025). Its stated aim is to overcome two practical constraints in PET AI—limited access to labeled data and insufficient computational resources—by combining projection-domain modeling with PET-specific geometric augmentation, structured mask constraints, and parameter-free lesion guidance. In a broader and earlier detector-architecture sense, closely related literature has also used “ALL-PET concepts” to describe fully RPC-based, non-scintillator PET systems; that usage is conceptually adjacent but technically distinct from the foundation-model framework named ALL-PET (1212.5551).

1. Problem setting and scope

ALL-PET is formulated against a specific mismatch between PET and contemporary foundation-model practice. The paper argues that PET data are difficult to collect and standardize because of privacy and ethics, high annotation cost, scanner and vendor variability, patient-to-patient physiological variability, and limited availability of raw projection data (Huang et al., 11 Sep 2025). It further contrasts its design with recent foundation models that use tens or hundreds of thousands to millions of images, very large parameter counts, and multi-node, multi-GPU training infrastructure. By contrast, ALL-PET is designed to work with 500 sinograms per task or condition and to train on a single NVIDIA RTX 3090 24 GB GPU.

A central claim is that image-domain PET foundation modeling is physically incomplete. The paper argues that reconstructed images already reflect an inverse problem and may discard or distort attenuation effects, angular and view dependence, and radial detector geometry that remain explicit in sinogram space. This motivates operating directly on $S(s,\theta)$, where the model can preserve physically meaningful line-integral structure and enforce geometric consistency more naturally than image-domain models.

The intended downstream scope is multi-task PET generation or translation in the projection domain. The paper lists low-dose PET reconstruction, attenuation correction or self-attenuation correction, delayed-frame prediction in dynamic PET, tracer conversion, and dual-tracer separation as target tasks, and the abstract additionally highlights tracer separation more generally (Huang et al., 11 Sep 2025). This suggests that ALL-PET is proposed not as a single-task reconstructor, but as a reusable PET generative backbone.

2. Projection-domain formulation and architectural design

At a systems level, ALL-PET is a latent diffusion model in sinogram space. Its pipeline begins with limited raw PET sinograms, augments and decomposes them using Radon-projected masks, encodes the result with a sinogram autoencoder, performs diffusion in latent space, and decodes synthetic or task-specific sinograms for optional FBP-based image reconstruction (Huang et al., 11 Sep 2025).

The core composite input is explicitly defined as

$$C_{\text{all}} = [C_{\text{pos}},\, S_{\text{full}},\, C_{\text{neg}}] \in \mathbb{R}^{3 \times H \times W},$$

with latent encoding

$z_0 = E(C_{\text{all}}) \in \mathbb{R}^d.$

The paper does not provide the exact latent dimensionality $d$, but it makes clear that diffusion is performed in latent rather than full-resolution sinogram space.

Three diffusion regimes are supported. DDPM and DDIM are used for unsupervised or generative settings, while DDBM is used for supervised source-to-target bridging tasks. In the DDPM formulation, the forward noising transition is written as

$$q(z_t\mid z_{t-1})=\mathcal N\!\left(z_t;\sqrt{\alpha_t}\,z_{t-1}, (1-\alpha_t)I\right), \qquad \alpha_t=1-\beta_t,$$

with closed-form marginal

$$q(z_t\mid z_0)=\mathcal N\!\left(z_t;\sqrt{\bar\alpha_t}\,z_0, (1-\bar\alpha_t)I\right),$$

and denoising objective

$$L_{\text{denoise}} = \mathbb E_{z_0,\epsilon,t} \left[ \left\| \epsilon - \epsilon_\theta\big( \sqrt{\bar\alpha_t}z_0+\sqrt{1-\bar\alpha_t}\epsilon,\ t \big) \right\|_2^2 \right].$$

The architectural emphasis is not merely on diffusion, but on constraining that diffusion by PET acquisition geometry. The denoising network is described as a time-aware neural denoiser $\epsilon_\theta$ in latent space, and conditioning enters through the three-channel masked/full/complement sinogram construction, through diffusion-mode selection, through prompt-like tracer or anatomy conditioning shown in figures, and through projected ROI guidance via transparent medical attention. The paper does not specify an exact text encoder or prompt embedding implementation for the prompt-like conditioning pathway.

Module	Representation	Function
Sinogram autoencoder	$C_{\text{all}} \rightarrow z_0$	Compresses 3-channel sinograms into latent space
Latent diffusion	DDPM / DDIM / DDBM	Supports unsupervised synthesis and supervised translation
PET-specific conditioning	Masks, prompts, ROI guidance	Injects geometric and task-specific structure
Optional reconstruction	Sinogram $\rightarrow$ FBP image	Enables image-space evaluation or interpretation

A notable misconception the paper explicitly rejects is that low-resource PET foundation modeling must simply miniaturize an image-domain FM. Its argument is instead that PET should be modeled where the acquisition physics are still explicit, namely the projection domain (Huang et al., 11 Sep 2025).

3. Radon mask augmentation, dynamic multi-mask generation, and geometric constraints

The most distinctive data-efficiency mechanism in ALL-PET is the Radon mask augmentation strategy, abbreviated RMAS or RMA. Rather than using generic image augmentation, the method generates randomized masks in image space, projects them to sinogram space with the Radon transform, and uses the resulting projection masks to decompose each real sinogram into visible and complementary parts (Huang et al., 11 Sep 2025).

The Radon transform is given as

$$C_{\text{all}} = [C_{\text{pos}},\, S_{\text{full}},\, C_{\text{neg}}] \in \mathbb{R}^{3 \times H \times W},$$0

For image-domain masks $$C_{\text{all}} = [C_{\text{pos}},\, S_{\text{full}},\, C_{\text{neg}}] \in \mathbb{R}^{3 \times H \times W},$$1, the projected discrete approximation is

$$C_{\text{all}} = [C_{\text{pos}},\, S_{\text{full}},\, C_{\text{neg}}] \in \mathbb{R}^{3 \times H \times W},$$2

and the sum over $$C_{\text{all}} = [C_{\text{pos}},\, S_{\text{full}},\, C_{\text{neg}}] \in \mathbb{R}^{3 \times H \times W},$$3 projected masks is

$$C_{\text{all}} = [C_{\text{pos}},\, S_{\text{full}},\, C_{\text{neg}}] \in \mathbb{R}^{3 \times H \times W},$$4

Normalization is then applied either by max-scaling

$$C_{\text{all}} = [C_{\text{pos}},\, S_{\text{full}},\, C_{\text{neg}}] \in \mathbb{R}^{3 \times H \times W},$$5

or clipping

$$C_{\text{all}} = [C_{\text{pos}},\, S_{\text{full}},\, C_{\text{neg}}] \in \mathbb{R}^{3 \times H \times W},$$6

The paper states that 64×64 image-domain mask blocks are used and that RMAS generates over 200,000 structurally diverse Radon masks from only 500 real sinograms. That diversity is then extended by dynamic multi-mask generation, or DMM, in which the number, size, and spatial distribution of non-overlapping masks vary from iteration to iteration. The paper’s argument is that DMM increases diversity without adding trainable parameters, because it modifies only the data pipeline.

This projected masking directly defines the positive and negative sinogram channels: $$C_{\text{all}} = [C_{\text{pos}},\, S_{\text{full}},\, C_{\text{neg}}] \in \mathbb{R}^{3 \times H \times W},$$7

$$C_{\text{all}} = [C_{\text{pos}},\, S_{\text{full}},\, C_{\text{neg}}] \in \mathbb{R}^{3 \times H \times W},$$8

These are not arbitrary auxiliary views; they are tied exactly to the same full sinogram by projection geometry. Their consistency is enforced through

$$C_{\text{all}} = [C_{\text{pos}},\, S_{\text{full}},\, C_{\text{neg}}] \in \mathbb{R}^{3 \times H \times W},$$9

The paper presents these positive and negative mask constraints as a zero-cost regularization mechanism. The claim is not that they reduce parameter count numerically layer by layer, but that they reduce the burden on the denoising backbone to infer projection geometry from scratch. This suggests a specific notion of “foundation” in ALL-PET: not massive scale, but structured inductive bias embedded in the data representation itself.

The paper also includes a PAC-Bayes analysis for augmentation. With probability at least $z_0 = E(C_{\text{all}}) \in \mathbb{R}^d.$0,

$z_0 = E(C_{\text{all}}) \in \mathbb{R}^d.$1

with looser bound

$z_0 = E(C_{\text{all}}) \in \mathbb{R}^d.$2

The paper reports lower generalization-error upper bounds after augmentation, with a reduction of about 0.027 for the shown confidence levels (Huang et al., 11 Sep 2025).

4. Transparent Medical Attention and clinician-adjustable guidance

Transparent medical attention, or TMA, is ALL-PET’s interpretability-oriented guidance mechanism. It is explicitly described as parameter-free and geometry-driven, in contrast to learned attention maps produced by trainable self-attention or cross-attention modules (Huang et al., 11 Sep 2025).

The procedure begins with a coarse image-space reconstruction $z_0 = E(C_{\text{all}}) \in \mathbb{R}^d.$3 obtained by FBP or OSEM. A binary lesion or ROI mask is then formed by thresholding: $z_0 = E(C_{\text{all}}) \in \mathbb{R}^d.$4 This mask is projected into sinogram space by

$z_0 = E(C_{\text{all}}) \in \mathbb{R}^d.$5

and used to gate feature maps: $z_0 = E(C_{\text{all}}) \in \mathbb{R}^d.$6

The paper states that TMA is intended to cover both hypermetabolic and hypometabolic abnormalities, and that lesion-focused attention maps are derived from coarse segmentation and projected into sinogram space for physically consistent guidance (Huang et al., 11 Sep 2025). It further states that clinician-defined ROI adjustments are supported, so that physicians can manually define or edit ROIs in image space and have those edits projected into sinogram space at inference without retraining. This is presented as a mechanism for flexible, interpretable, and task-adaptive emphasis aligned with PET acquisition physics.

An important distinction follows from this design. In ALL-PET, “attention” does not primarily mean a learned transformer operation. It means deterministic ROI-weighting constructed from coarse reconstruction, thresholding, projection geometry, and elementwise gating. The paper’s terminology is therefore closer to transparent feature modulation than to opaque learned relevance estimation.

5. Data regime, tasks, and reported empirical behavior

The empirical setup is deliberately small-scale in real data and expansive in augmentation. Data were acquired on a DigitMI 930 PET/CT scanner, with AFOV 30.6 cm and ring diameter 81 cm, from 14 patients scanned over 4 to 8 beds, with full sampling scan time from 45 s to 3 min per bed (Huang et al., 11 Sep 2025). The paper states that 500 sinograms were randomly sampled for training and 100 for testing. For synthetic tasks, it uses 500 sinograms per tracer or anatomical region; for supervised downstream tasks, 500 paired sinograms for training and 100 paired sinograms for testing. The paper does not explicitly define a separate validation split.

The synthetic training spans three tracers—$z_0 = E(C_{\text{all}}) \in \mathbb{R}^d.$7F-FDG, $z_0 = E(C_{\text{all}}) \in \mathbb{R}^d.$8F-DOPA, and $z_0 = E(C_{\text{all}}) \in \mathbb{R}^d.$9Ga-PSMA—and anatomical regions including brain, trunk, chest, and abdomen. Downstream paired tasks include low-dose versus normal-dose, pre- versus post-attenuation correction, first-scan versus delayed-scan, dual-tracer mixture versus separated tracers, and tracer-to-tracer conversion. The paper also notes extension potential to cross-modality integration with CT, although the detailed experiments in the provided text remain PET-centric.

Several implementation details are explicit. Example sampled data shape is $d$0, and the main memory-saving design choices are latent diffusion instead of pixel-space diffusion, geometry-driven augmentation instead of large-scale data collection, parameter-free TMA instead of learned attention blocks, and positive or negative constraints as zero-cost regularizers. The abstract and details jointly state that ALL-PET operates efficiently with memory use under 24GB and is trainable on a single NVIDIA RTX 3090 24 GB GPU (Huang et al., 11 Sep 2025).

The paper’s empirical claims are high-level rather than metric-dense. It reports high-quality sinogram generation using only 500 samples, performance comparable to models trained on larger datasets, and generalization across low-dose reconstruction, attenuation correction, delayed-frame prediction, and tracer separation. It also reports that over 200,000 structurally diverse training samples are generated from the limited real set. By design, the paper’s evidence emphasizes feasibility under low-resource constraints rather than leaderboard-style metric saturation.

At the same time, the manuscript leaves several technical details under-specified. The exact latent dimensionality is not reported, the precise prompt-conditioning mechanism is not detailed, preprocessing beyond sinogram size, coarse FBP or OSEM reconstruction, and mask normalization is limited, and several DDBM-related expressions are described as typeset imperfectly in the provided summary. This suggests that ALL-PET is best understood as a physically structured PET FM proposal with strong systems-level motivation, rather than as a fully standardized benchmark package.

6. Relation to broader PET foundation modeling and alternate uses of the term

ALL-PET occupies a distinctive place in recent PET AI because it is projection-domain native. Most recent PET foundation-model or PET-scale model work remains image-domain. SegAnyPET, for example, defines a modality-specific 3D foundation model for universal promptable segmentation from whole-body $d$1F-FDG PET images using the PETS-5k dataset of 5,731 3D scans and a cross prompting confident learning strategy, but it operates on reconstructed PET volumes rather than raw sinograms (Zhang et al., 20 Feb 2025). PETITE, in turn, studies parameter-efficient fine-tuning for multi-scanner PET-to-PET reconstruction and shows near-full-fine-tuning performance with less than 1% trainable parameters, but it is likewise image-domain and scanner-adaptation oriented (Kim et al., 2024). PET-CM applies a consistency-model diffusion framework to synthesize full-dose whole-body PET from low-dose PET, reporting 62 seconds generation time per patient and a 12× speedup relative to a diffusion-based baseline, but again in reconstructed-image space (Pan et al., 2023).

This contrast is important because ALL-PET’s central claim is not simply low-shot PET generation, but low-shot PET generation where acquisition geometry remains explicit. That position is conceptually aligned with the broader industry argument that richer upstream PET data representations, including raw or standardized list-mode-like forms, may enable more flexible AI pipelines than fixed reconstructed images (Sitek et al., 2021). ALL-PET can therefore be read as one concrete response to the long-standing call for PET AI systems that engage more directly with acquisition-domain information.

The term “ALL-PET” also has a broader historical resonance in detector-concept literature. The multigap RPC study on PET detector optimization is described as directly relevant to ALL-PET concepts in the sense of PET systems that replace scintillator crystals with RPC-based or other non-scintillator detector architectures (1212.5551). In that literature, the phrase points toward an all-RPC or fully non-scintillator PET system, whereas in (Huang et al., 11 Sep 2025) it names a PET foundation model in the projection domain. This suggests that the term now straddles two distinct usages: one at the detector-architecture level and one at the PET-AI model level.

Taken together, these strands define ALL-PET less as a settled category than as a convergence point around PET-wide generality. In the detector sense, it implies PET systems built from alternative sensing architectures; in the model sense, it implies a reusable PET generative backbone that is low-shot, low-resource, and physically grounded in projection data. The latter is the specific and formal meaning established by the paper titled "ALL-PET: A Low-resource and Low-shot PET Foundation Model in the Projection Domain" (Huang et al., 11 Sep 2025).