A Convolutional Neural Network Based Iterative Framework for Plant PET Image Denoising

Plant positron emission tomography (PET) images usually suffer from noise because of positron spillover effect. Hence, it is worth performing image denoising. In this paper, we propose a convolutional neural network (CNN) based iterative framework and unfold the CNN-based iterative framework into a deep network, equipped with an unsupervised learning strategy, namely regularized Noise2noise. The iterative framework mathematically solves the denoising problem by half quadratic splitting and gradient descent method, and benefits from the CNN capacity and flexibility of exploiting image characteristics. Furthermore, the regularized Noise2noise provides a constraint that considers the difference between the ground truth corresponding to the input and to the label. The iterative framework was tested on dynamic plant PET images and compare with three popular denoising methods. Experimental results show that the iterative framework is effective in boosting the denoising performance. The iterative framework achieves an average signals to noise ratio of 34.93 and an average contrast to noise ratio of 62.51, outperforming all compared methods.