Deep-learning-based Partial Volume Correction in 99mTc-TRODAT-1 SPECT for Parkinson's Disease: A Preliminary Study on Clinical Translation.

Abstract

^99mTc-TRODAT-1 SPECT is effective for the early detection of Parkinson's disease (PD). However, SPECT images suffer from severe partial volume effect, which impairs tissue boundary clarity and subsequent quantification accuracy. This work proposes an anatomical prior- and segmentation-free deep learning (DL)-based partial volume correction (PVC) method using an attentionbased conditional generative adversarial network (Att-cGAN) for ^99mTc-TRODAT-1 SPECT. A population of 454 digital brain phantoms modelling anatomical and ^99mTc-TRODAT activity variations in different PD categories are used to generate realistic SPECT projections using the SIMIND Monte Carlo code, and then reconstructed using ordered subset expectation maximization algorithm. The dataset is split into 320, 44 and 90 used for training, validation, and testing. Att-cGAN, cGAN and U-Net are implemented based on simulated data, then directly tested on 100 retrospectively collected clinical ^99mTc-TRODAT data, with same acquisition and reconstruction parameters as in simulations. Non-DL PVC methods of Van-Cittert and iterative Yang are implemented for comparison. Physical and clinical metrics, as well as a no-gold standard technique (NGST) are applied to evaluate different PVC methods in the absence of clinical ground truth. Att-cGAN yields superior PVC performance in simulations as compared to other methods in physical and clinical evaluations. NGST assessment is generally consistent with the clinical metric evaluation. For the clinical study, Att-cGAN also obtains better NGST result than others striatal compartments can be discriminated on DLbased processed images. DL-PVC method is feasible for clinical PD SPECT using highly realistic simulated data.