PET Images Enhancement Using Deep Training of Reconstructed Images with Bayesian Penalized Likelihood Algorithm

Abstract

Purpose

To adopt the merits of the Bayesian Penalized Likelihood (BPL) reconstruction algorithm (incl. improved contrast recovery), a deep learning ResNet model was trained to estimate BPL-like images using the non-attenuation, non-scatter corrected PET images (PET-nonAC) as inputs.

Methods

Images of 112 patients were used for model training (79 patients), validation (13 patients) and testing (20 patients). The ResNet model used PET-nonAC images as input and predicted corresponding BPL-like images. The model performance regarding image quality was evaluated using metrics such as contrast-to-noise ratio (CNR).

Results

The CNR of the reference BPL images was 2.40, while estimated BPL-like images using the deep learning model have a CNR value of 2.42 indicative of comparable performance.

Conclusion

The estimated BPL-like images of the deep learning model offer comparable quality to the reference BPL images especially regarding the CNR metric. This deep learning model can be used to improve the image quality PET-nonAC by adopting the characteristics of the BPL images.