Standard-dose PET(SPET) Images Synthesizing

Abstract

Positron emission tomography (PET) is a functional imaging modality that uses a radioactive tracer to visualize and quantify metabolic processes in the human body. It has demonstrated considerable clinical value in oncology neuropsychiatry and cardiology showing substantial promise for advancing cancer diagnosis and management cardiac care and surgery as well as neurological and psychiatric applications. However repeated imaging poses health risks due to cumulative radiation exposure. To address this critical issue there have been attempts to lower the injected activity to get low-dose PET (LPET) images. Because LPET involves less radiotracer accumulation than standard-dose PET (SPET) the reconstructions are noisier and can compromise diagnostic utility. To restore image quality recent work employs deep-learning methods to denoise LPET and synthesize images that closely approximate SPET quality. This thesis presents new deep learning methods for PET synthesis. First to counter the scarcity of paired LPET–SPET data we propose a data-augmentation pipeline that synthesizes LPET from existing SPET scans. Second we introduce a time-controlled model that encodes noise level as a timestep variable. Trained only on the lowest-dose data the model generalizes to unseen dose conditions (e.g. 50 10 1 ). Both methods were trained on a public dataset called ultra-low-to-high PET dataset. A subset containing PET images scanned by a Siemens scanner was selected with each subject having seven different dose levels. The augmentation strategy boosts performance over training with smaller datasets and works with several SOTA synthesis networks. The time-controlled approach outperforms strong baselines and remains robust when tested on dose levels withheld during training highlighting flexibility and wide generalization.