Application of trained Deep BCD-Net to iterative low-count PET image reconstruction

Image reconstruction in low-count PET is challenging because gammas from natural radioactivity in Lu-based crystals cause high random fractions that lower the measurement signal-to-noise-ratio (SNR). In model-based image reconstruction, using more iterations of an unregularized method may increase the noise, so incorporating regularization into the image reconstruction is desirable to control the noise. New regularization methods based on learned convolutional operators are emerging in iterative reconstruction. We apply the newly developed block coordinate descent network (BCD-Net) to PET reconstruction by modifying the image reconstruction module to incorporate PET physics. Using the XCAT phantom we simulated the low true coincidence count-rates with high random fractions typical for Y-90 PET patient imaging after Y-90 microsphere radioembolization. We trained a 10 layer BCD-Net where each layer has 200 convolutional filters that encode/decode an input image. Numerical results show that deep BCD-Net significantly improves PET reconstruction performance compared to iterative image reconstruction using non-trained regularizers (total variation (TV) and non-local means (NLM)). We selected the regularization parameter for each method to obtain the highest contrast to noise ratio (CNR). BCD-Net improved activity recovery for a hot sphere significantly and reduced noise, whereas non-trained regularizers had a tradeoff between noise and quantification. BCD-Net improved CNR and activity recovery by 96.6% (150.0%) and 41.5% (35.9%) compared to TV (NLM) regularized reconstruction.