Resolution and quality enhancement of SPECT cerebral blood flow images using Pix2pix deep learning

Cerebral blood flow (CBF) imaging can be performed using SPECT with ¹²³I-IMP; however, its spatial resolution and image quality are inferior to those of PET-CBF imaging using labeled water. This study aimed to enhance the resolution and image quality of SPECT-CBF images using the pix2pix machine learning framework.

Seventy-three patients with suspected cerebral ischemia underwent CBF imaging using SPECT (Symbia, with ¹²³I-IMP) and PET (mCT and Vision, with O-15-labeled gas). Image reconstruction was performed using OSEM for PET and Flash3D for SPECT. The SPECT and PET images were coregistered using SPM12, and a pix2pix model was trained using SPECT-CBF images as input and PET-CBF images as the target, with 43 cases used for training and 15 for testing. P2P-SPECT-CBF images were then generated for 15 cases for validation. Visual assessment on a 5-point scale, structural similarity index measure (SSIM), and region-of-interest (ROI)-based quantitative analysis were performed to evaluate image similarity and accuracy.

The P2P-SPECT-CBF images demonstrated improved visual similarity to PET images, with an average score of 4.2 and 3.5 in open and blind assessments, respectively. The SSIM value of conventional SPECT images compared to PET was 0.79, while that of P2P images was 0.86 and those were significantly different, indicating enhanced structural similarity. In ROI analysis, the correlation between SPECT and PET CBF values was y = 0.13 + 0.63x, r = 0.77 (p < 0.01). The correlation between P2P-SPECT and PET was y = 0.10 + 0.65x, r = 0.78 (p < 0.01), and between P2P-SPECT and SPECT, the relationship was y = 0.01 + 0.89x, r = 0.86 (p < 0.01).

The proposed method generated P2P-SPECT-CBF images with image contrast closely resembling that of PET-CBF, while preserving the quantitative properties of SPECT-CBF.