Combining GAN Loss with Pre-trained Model for Semi-supervised SPECT Reconstruction

The radioactive tracers used in Single-Photon Emission computerized Tomography (SPECT) scans can cause radiation exposure to the human body. Therefore, low-dose SPECT has attracted widespread attention in nuclear medicine imaging. Under low-dose imaging conditions, projection data is heavily contaminated by severe noise. There has been a significant amount of studies that explore fully supervised deep learning reconstruction methods to suppress image noise. The quality of images obtained by fully supervised methods depend on the quantity and quality of the labels. However, it is challenging to obtain the normal-dose images with supervised labels in clinical practice. To overcome the challenge, we propose a pre-trained mean teacher method with GAN loss to achieve low-dose SPECT reconstruction. The proposed method introduces a Swin-Conv-Unet-based pre-trained model into the mean teacher model to enhance the reliability of unlabeled training data. The teacher model supervises the student model through consistency regularization; the pre-trained model is trained with a small amount of labeled data and enhances the supervision reliability through GAN loss. Numerical experiments validate the performance of the proposed method in noise suppression and feature preservation. When compared with the mean teacher method, the SSIM of the reconstructed images is increased by 2%, the RMSE is reduced by 9%, and the PSNR is increased by 0.77 dB. The dataset is generated by the SIMIND simulation software using XCAT digital phantoms.