Deep Transformer Based Direct Attenuation Correction of Myocardial Prefusion SPECT Images

Introduction: Convolutional neural network (CNN)-based algorithms, such as U-Net and GAN, suffer from a number of limitations in explicit long-range and global relation modeling, owing to convolution operations locality. This might result in poor performance in patients with large variability. In addition, down and up sampling processes in these algorithms result in loss of spatial information. Transformers introduced in Natural Language Processing (NLP) domain have been recently proposed to tackle these issues in the image domain. Because of limited resolution of SPECT images and high variability resulting from the use of different cameras on various patients, we propose a Deep Transformer for direct attenuation correction of cardiac SPECT images.

Methods: We enrolled 85 patients who underwent 99mTc-sestamibi myocardial perfusion SPECT (MPI-SPECT) imaging on two different SPECT/CT cameras. SPECT images were obtained using a dual-headed gamma camera with body auto-contour form 135° (RAO) to -45° (LAO) in 180° orbit, matrix size 64×64, 32 steps with 30 seconds per step and 16-bin gating. The photopeak was set to 20% energy window around 140 keV. Image reconstruction was performed using OSEM with 4 iterations and 4 subsets and a post-reconstruction Gaussian filter (FWHM= 12 mm). Low-dose CT scanning was performed for attenuation correction. Data were split to 80% train/validation and 20% test set. Swin Transformers-based network consisting of encoder, decoder, skip connection, and bottleneck, as pure U-Shape Transformer were implemented. The images were split to patches (extracted from 2D slices) used as a token for transformers. Non-attenuation corrected images were used as input of the network which attempted to generate CT-based attenuation corrected images using 2D training process. Images were evaluated quantitatively using different metrics.

Results: Overall, the transformer-based algorithm showed good agreement with CT-based attenuation corrected images as ground truth. The proposed models resulted in percent relative error (RE%) of 2.1±4.1%, structural similarity index (SSIM) of 0.99±0.01 and peak signal to noise ratio (PSNR) of 36.2±2.2, respectively.

Conclusions: This study showed that the transformer-based deep neural network can achieve accurate attenuation correction in MPI-SPECT, which is quantitatively comparable to CT-based attenuation correction. In spite of the low resolution of SPECT images and the high variability, transformers-based algorithms produced high quality attenuated corrected SPECT images suitable for application on standalone dedicated cardiac SPECT cameras.

• Download figure
• Open in new tab
• Download powerpoint