TY - JOUR T1 - Wavelet-based PET image restoration using an artificial neural network and maximum likelihood-expectation maximization algorithm JF - Journal of Nuclear Medicine JO - J Nucl Med SP - 1481 LP - 1481 VL - 50 IS - supplement 2 AU - Jih-Shian Lee AU - Kuan-Hao Su AU - Ren-Shyan Liu AU - Chen Jyh-Cheng Y1 - 2009/05/01 UR - http://jnm.snmjournals.org/content/50/supplement_2/1481.abstract N2 - 1481 Objectives The purpose of this study was to obtain a shift-variant point spread function (SVPSF) using an artificial neural network (ANN) and to restore the PET images with the SVPSF using maximum likelihood-expectation maximization (MLEM). Methods A Mini-Deluxe phantom was used to acquire images from the microPET® R4 and a high-resolution optical scanner for ANN training. First, wavelet de-noising technique was used to reduce the noise of PET images. Then, the high-resolution scanner images and the PET images were used as inputs and outputs to obtain the SVPSF using a two-layer-architecture ANN. To restore image, we acquired the PET image followed by wavelet de-noising. Second, the MLEM was used to restore the PET image with the ANN corrected system matrix. We used the Mini-Deluxe and multi-line source phantom to calculate root mean square error (RMSE), contrast recovery coefficient (CRC), and resolution for evaluation. Results The RMSE between the true and original PET image (without image restoration) is 0.21 and is 0.19 between the true and restored image. The CRC of the original PET image is 0.69 and is 0.79 for the restored image. In resolution study, the full width at half maximum (FWHM) of the original PET image is 1.65, 1.69, 2.35, and 2.69 mm from center to edge, and is 1.56,1.60, 2.08, and 2.26 mm after restoration. Conclusions The resolution of the restored image is better and more uniformly distributed than the original one. The CRC is also better. The results suggested that PET images can be restored by the proposed method for improving the image quality. ER -