🔬 physics

Two-stage deep learning framework for the restoration of incomplete-ring PET images

This paper presents a two-stage deep learning framework that effectively restores high-quality PET images from data with over 50% missing coincidences by combining a projection-domain Attention U-Net for sinogram completion with a cascaded U-Net and warm-start diffusion model for image refinement, achieving near-complete anatomical fidelity without requiring time-of-flight information.

Original authors: Yeqi Fang, Rong Zhou

Published 2026-05-26

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Yeqi Fang, Rong Zhou

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to solve a massive, 3D jigsaw puzzle of a human brain, but someone has ripped out nearly half the pieces. In the world of medical imaging, this is what happens when a PET scanner (a machine that takes pictures of how your body uses energy) has missing parts.

Usually, these scanners are complete rings of detectors, like a perfect hula hoop. But sometimes, due to broken parts, cost-cutting, or specific medical needs (like scanning a patient's breast or keeping a patient from feeling claustrophobic), the ring is broken. It might be missing 50% of its detectors. When you try to build the picture with only half the data, the result is usually a blurry, distorted mess full of "ghost" lines and missing details.

This paper introduces a two-stage "AI repair crew" that can fix these broken pictures, even when half the data is missing. Here is how they did it, explained simply:

The Problem: The Broken Ring

Think of the scanner as a camera that takes a photo from every angle around a patient. If the camera is missing half its lenses, the photo comes out with huge gaps. Traditional math methods try to guess what's in the gaps, but they often get it wrong, creating streaks and blurs that look like static on an old TV.

The Solution: A Two-Step Repair Process

The researchers built a system that acts like a master restorer, working in two distinct stages:

Stage 1: The "Sinogram" Detective (Filling the Gaps)

Before the computer makes the final brain image, it first creates a raw data map called a sinogram. You can think of a sinogram as the "blueprint" or the "negative" of the photo.

The Challenge: The blueprint has huge holes in it because the scanner was broken.
The Fix: The team used a special AI called an Attention U-Net. Imagine this AI as a detective who looks at the edges of the missing puzzle pieces.
- Instead of just guessing randomly, the AI looks at the slices next to the missing piece and the slices above and below it (like looking at the pages before and after a torn page in a book).
- It uses an "Attention" mechanism, which is like a spotlight. It tells the AI, "Hey, focus on these specific lines and shapes; ignore the noise."
- This AI successfully fills in the missing blueprint data, turning a broken map into a complete one.

Stage 2: The "Refinement" Artist (Polishing the Image)

Once the blueprint is fixed, the computer builds the first draft of the brain image using standard math (called OSEM). But this first draft is still a bit rough, like a sketch that needs shading.

The Fix: The team uses a second AI, which combines a 3D U-Net and a Diffusion Model.
- Think of the Diffusion Model as a "denoising" tool. Usually, these tools start with a blank canvas of pure static (random noise) and slowly turn it into a picture.
- The Innovation: Instead of starting from scratch, this AI starts with the "rough sketch" from the first stage. It treats the rough sketch as a "warm start."
- It then acts like a sculptor, chipping away the errors and adding back the fine details (like the texture of the brain's surface) that the first draft missed. It learns exactly what is missing from the rough sketch and adds it back in.

The Results: A Near-Perfect Restoration

The team tested this on 613 simulated brain scans (including healthy brains, brains with Alzheimer's, and brains with mild cognitive issues).

The Damage: They simulated scanners missing about 50% of their data (double the damage previous methods could handle).
The Outcome: The AI restored the images so well that they looked almost identical to the original, perfect scans.
- They measured the quality using two scores: PSNR (how close the colors and brightness are) and SSIM (how similar the shapes and structures are).
- The results were incredibly high (SSIM of 0.99 out of 1.0), meaning the restored images preserved the brain's anatomy almost perfectly.

What This Means (and What It Doesn't)

What it does: It proves that with this specific two-step AI method, we can take data from a broken, incomplete scanner and turn it into a high-quality, diagnostic-ready image. It works even without special "time-of-flight" hardware (which usually helps fix these problems).
What it doesn't do (yet): The paper is careful to say this was tested on simulated data (computer-generated models based on real patient data), not on actual patients in a hospital yet. The authors note that real-world scanners have more complex "glitches" than their simulation, and they haven't tested this on tumors or other organs yet.

In short: The researchers built a digital "magic eraser" and "smart painter" that can look at a half-broken medical scan, figure out what the missing half should look like based on the surrounding clues, and paint a perfect picture, effectively saving the diagnostic value of a broken machine.

Technical Summary: Two-Stage Deep Learning Framework for Incomplete-Ring PET Restoration

Problem Statement
Positron Emission Tomography (PET) typically relies on complete 360° detector rings to ensure full angular coverage and reliable data collection. However, incomplete-ring PET scanners are increasingly utilized due to hardware failures, cost constraints, or specific clinical needs (e.g., breast imaging, in-beam therapy monitoring, and open-configuration brain scanners). These systems suffer from reduced data completeness and geometric inconsistencies, rendering standard reconstruction algorithms (such as Filtered Back-Projection or MLEM) ineffective due to pronounced streak artifacts and resolution non-uniformity. While Time-of-Flight (TOF) capabilities offer partial mitigation, they cannot fully compensate for severe angular deficits. Previous deep learning approaches have largely addressed mild data loss (up to ~25% coincidence loss) or relied on unstable Generative Adversarial Networks (GANs) and slow explicit likelihood models. There is a lack of robust methods capable of restoring high-quality images from data with approximately 50% missing coincidences without requiring TOF information.

Methodology
The authors propose a two-stage deep learning framework designed to restore high-quality PET images from incomplete-ring data. The pipeline operates without incorporating TOF information and proceeds as follows:

Stage 1: Sinogram Restoration (Projection Domain)
- Input Strategy: The system utilizes a five-channel input strategy. The central channel represents the current sinogram slice, while the other four channels incorporate spatial context from neighboring slices ( $j-1, j+1$ ) and temporal context from adjacent sinogram periods ( $j-R, j+R$ ).
- Architecture: An Attention U-Net is employed to predict the missing sections of the sinogram. Unlike standard U-Nets, this architecture integrates Attention Gating (AG) modules in the skip connections. These modules adaptively highlight significant structural features and suppress irrelevant regions, allowing the model to focus on structural continuity around missing angular data.
- Training: The model is trained using the Adam optimizer with mixed precision and gradient scaling. It learns to map incomplete sinograms to complete ones by leveraging the spatial and temporal correlations within the sinogram tensor.
Stage 2: Image Refinement (Image Domain)
- Initial Reconstruction: The completed sinograms from Stage 1 are reconstructed into preliminary images using the Ordered Subsets Expectation Maximization (OSEM) algorithm.
- Refinement Architecture: A cascaded dual-model architecture refines the OSEM output.
  - A 3D U-Net generates a coarse prediction ( $I_P$ ) from the OSEM image.
  - A Diffusion Model (specifically a Resfusion-based approach) performs image refinement. Crucially, this is a "warm-start" diffusion process: instead of starting from pure Gaussian noise, the reverse diffusion process begins from a noisy version of the U-Net's coarse prediction ( $I_P$ ).
  - The diffusion denoiser (a separate U-Net backbone) learns to predict the residual ( $x_0 - I_P$ ), which is added to the coarse prediction to produce the final high-fidelity image.

Dataset and Experimental Setup

Data Source: The study utilizes a dataset of 613 simulated brain volumes derived from the Alzheimer's Disease Neuroimaging Initiative (ADNI), comprising 196 healthy controls, 217 Alzheimer's disease (AD) subjects, and 200 Mild Cognitive Impairment (MCI) subjects.
Simulation: A custom Monte Carlo simulation framework (SimSET) was used to model a PET scanner with 42 axial rings and 182 detectors per ring. Incomplete-ring geometries were simulated by masking specific detector rings or angular sectors, resulting in approximately 50% coincidence loss (double the severity of previous CNN-based benchmarks).
Validation: The study employs 6-fold cross-validation and evaluates performance under various angular-loss patterns (e.g., contiguous gaps vs. interleaved gaps).

Key Results
The framework demonstrates robust performance in recovering anatomical structures and tracer distributions under severe data loss:

Quantitative Metrics: The final refined images achieved a PSNR of 38.18–38.59 dB and an SSIM of 0.9904–0.9925.
Ablation Studies:
- The first-stage Attention U-Net is the most critical component; removing it causes SSIM to drop significantly (from 0.9925 to 0.9293 when using only OSEM + 3D U-Net).
- The Attention mechanism proved superior to standard U-Nets, UNETR, and TransUNet, yielding a 1.24 dB PSNR increase and 0.052 SSIM increase over the baseline U-Net.
- The second-stage diffusion refinement provided a significant boost, improving SSIM from 0.9860 (Attention U-Net + OSEM + 3D U-Net) to 0.9925.
Visual Quality: The model successfully preserved key anatomical boundaries (e.g., cerebral cortex, basal ganglia) and high-uptake regions, effectively mitigating the streak artifacts typically associated with incomplete angular sampling.

Significance and Claims
The authors claim this work provides the first systematic benchmark for incomplete-ring PET reconstruction at the severity of ~50% coincidence loss, a level previously unaddressed by CNN-based methods (which were limited to ~25% loss).

Robustness: The framework achieves near-complete anatomical fidelity without requiring TOF information, making it applicable to a wider range of hardware configurations.
Scalability: Unlike unsupervised methods that require re-optimization for every scan, this supervised framework offers high throughput (approx. 18.6 seconds per volume on an NVIDIA RTX 4090).
Clinical Relevance: By validating on a large, diverse dataset (613 volumes) covering healthy, AD, and MCI populations, the study establishes a more rigorous statistical basis than previous works limited to small phantom sets (e.g., 20 phantoms).

Limitations and Future Directions
The paper modestly acknowledges several limitations:

Simulation Fidelity: The data is based on idealized Monte Carlo simulations (SimSET) and does not yet reflect the full range of crystal-to-crystal variations found in commercial scanners.
Anatomical Scope: Validation is currently restricted to brain PET; performance on other organs or aggressive pathological uptake patterns (e.g., focal tumors) remains unknown.
Dimensionality: The network processes 2D slices augmented with neighbor information, potentially sacrificing some genuine 3D context.
Future Work: The authors prioritize extending validation to broader pathological patterns (specifically for in-beam therapy monitoring), migrating to GATE-based simulations for higher credibility, and exploring hybrid 3D approaches (recurrent/transformer attention) to reduce memory costs while preserving inter-slice relationships.