Generating Artificial PET Scans from Low Dose CT Scans with an Adapted Dual Diffusion Implicit Bridges (DDIBs) model

Francesca Marogna, Martijn van Leeuwen, Fabian Hoitsma, Görkem Saygili, Sharon Ong

Research output: Chapter in Book/Report/Conference proceedingConference contributionScientificpeer-review

Abstract

Positron emission tomography (PET) scanning is an important tool to visualize abnormal metabolic activity, which is essential for detecting diseases such as cancer. PET scans are both expensive to acquire and uncomfortable for patients. A more cost-effective and less invasive imaging modality is Computed Tomography (CT) scanning. This paper compares the use of two generative models; a diffusion model and a Conditional Generative Adversarial Network (cGAN) to generate artificial PET scans from CT scans. A dataset from the Cancer Imaging Archive of 1014 whole-body PET/CT scans, with 501 cases of patients diagnosed with cancer was used. The pixels corresponding to cancer tissue in this dataset were annotated. 2D crops of 128 × 128 pixels were extracted in the axial plane, around the center of each cancerous tissue. The data was split subject-wise to create 23K training, 3K validation, and 2.5K test images. In this paper, we modified a Dual Diffusion Implicit Bridges (DDIBs) model to generate artificial PET images from CT images. DDIBs independently train two diffusion models; the source model obtains latent encodings from the CT images and the target model decodes these encodings to construct the PET images. We evaluated the performance of the DDIB model with the Structural Similarity Index Measure (SSIM) and Peak Signal-to-Noise Ratio (PSNR). We validated the adaptation of the DDIB model by reconstructing the original CT scans with the target diffusion model. How ever, our results showed that the DDIB model could not create high quality PET scans. To our knowledge, our work is the first adapt the DDIB model to generate PET scans from low-dose CT scans. Successful creation of synthetic PET scans can help clinicians decide whether an expensive PET scan is necessary based on an already-made CT scan.
Original languageEnglish
Title of host publication35th Benelux Conference on Artificial Intelligence and the 32th Belgian Dutch Conference on Machine Learning
Publication statusPublished - 19 Nov 2024

Fingerprint

Dive into the research topics of 'Generating Artificial PET Scans from Low Dose CT Scans with an Adapted Dual Diffusion Implicit Bridges (DDIBs) model'. Together they form a unique fingerprint.

Cite this