Diffusion-Based Sinogram Interpolation for Limited Angle PET

Abstract

Modern Positron Emission Tomography (PET) is increasingly adopting flexible and applicationdriven detector designs that deviate from traditional cylindrical configurations, such as walkthrough geometries or partial rings. These designs, while advantageous for patient comfort and system integration, lead to incomplete angular coverage and consequently, severely undersampled sinograms. Such limitations result in artifacts and degraded image quality when conventional reconstruction algorithms such as Maximum Likelihood Expectation Maximization (MLEM) are applied. To address this, we propose a data-driven framework that uses conditional diffusion models to recover the missing sinogram regions for improved image reconstruction. Our method leverages the pretrained Stable Diffusion (SD) model, which we fine-tune to predict missing sinogram bins based on limited-angle (LA) inputs. Training is conducted on 2D sinograms derived via the Radon transform from a publicly available PET/CT dataset. The LA regions are simulated by randomly omitting consecutive angular bins. To integrate conditional inputs, we insert convolutional blocks at multiple scales before and after the original SD network, allowing the model to effectively incorporate the LA sinogram into its prediction. While these blocks are being trained, the UNet blocks of the SD are simultaneously fine-tuned to adapt the network, originally trained on natural images, to the domain of PET sinograms, thereby mitigating errors caused by domain shift. The model learns a prior over fully sampled sinograms while using the available partial measurements as additional information during the denoising process. When combined with the MLEM reconstruction, our approach yields an average improvement of approximately 8 dB in peak signal-to-noise ratio (PSNR), highlighting the potential of diffusion-based generative approaches for robust PET imaging in non-standard acquisition settings.

Authors 3

1. Julian Thull Aachen Institute of Imaging & Computer Vision

   RWTH Aachen University

   Affiliation as printed

   RWTH Aachen University, Institute of Imaging and Computer Vision, Aachen,North Rhine-Westphalia,Germany

2. Johannes Stegmaier Aachen Institute of Imaging & Computer Vision

   RWTH Aachen University

   Affiliation as printed

   RWTH Aachen University, Institute of Imaging and Computer Vision, Aachen,North Rhine-Westphalia,Germany

3. Volkmar Schulz Aachen Institute of Imaging & Computer Vision

   RWTH Aachen University

   Affiliation as printed

   RWTH Aachen University, Institute of Imaging and Computer Vision, Aachen,North Rhine-Westphalia,Germany