Speaker
Description
Positron emission tomography (PET) is a molecular imaging technique that detects gamma photons from an injected radiotracer to map biological activity in the body. Accurate attenuation correction provided by computed tomography (CT) maps is important for PET to increase image quality, but standalone PET scanners don’t have a co-registered CT map which limits the correction quality and decreases clinical interpretability. This project’s goal is to develop a CT free attenuation correction workflow for dedicated brain PET by synthesizing pseudo-CT images from the PET system and then converting them into attenuation maps (μ-maps). A deep learning based Dual-Stage Generative Adversarial Network (DSG-GAN) model was trained on paired brain PET/CT data to learn the PET to CT translation, after which the synthetic CT outputs were converted to μ-maps using tissue segmentation and HU-to-attenuation conversion with post-processing steps (threshold tuning, bilinear scaling, and intensity normalization) to improve anatomical quality. A template-based pipeline was used to support hybrid solutions, including affine PET-to-PET registration to template space and deformation of template μ-maps using the same transform. Quantitative evaluation tools, including Structural Similarity Index (SSIM), Peak Signal-to-Noise Ratio (PSNR), and voxel-wise error mapping. Initial affine registration tests were used to evaluate the μ-maps. Preliminary work on a reduced subset produced promising synthetic CT-derived μ-maps, with next steps including training on the deep learning model on the full PET/CT dataset and developing comprehensive benchmarking of attenuation-corrected reconstructions.
| Keyword-1 | Positron Emission Tomography |
|---|---|
| Keyword-2 | Deep Learning Model |
| Keyword-3 | μ-map |