Improve Image Quality of High Spatial Resolution SPECT by Deep Image Prior

Abstract

Single photon emission computed tomography (SPECT) can visualize the functional information of the body. We are developing SPECT system with high spatial resolution of approximately 3.5 mm, like positron emission tomography (PET). However, image acquired by our SPECT system is noisy due to insufficient sensitivity. We aimed to improve the image quality of SPECT by using deep image prior (DIP), which does not require prior learning and can recover images using only the target image. To demonstrate usefulness of DIP for our SPECT system, simulation using a numerical phantom and measurement with physical phantom using a high-resolution SPECT system were performed. Projection image data processed by DIP were reconstructed using ordered subset expectation maximization (OS-EM) method. The reconstructed SPECT images with and without DIP were compared. In the simulation, DIP processing improved peak signal-to-noise ratio (PSNR) from 32.26 dB to 41.54 dB and structural similarity (SSIM) from 0.860 to 0.9439. Visual evaluation also demonstrated that the noise in the reconstructed image was reduced, and the structure of the phantom became clearer. In an experiment using a physical phantom, DIP also significantly suppressed the noise and clearly depicted the structure of the brain phantom. These results suggest that DIP is effective in reducing noise and improving image quality for high spatial resolution SPECT.