Feature Oriented Deep Convolutional Neural Network for PET Image Denoising

Deep convolutional neural networks have demonstrated superior performance in natural image denoising. Trainable network weights have been typically optimized by minimizing a loss function that computes pixel-wise discrepancies between the noisy image and the clean target image. In this study, we investigate an alternative solution that utilizes more prior knowledge to highlight the features of interest by modifying their contributions to the global loss function. We propose a feature-oriented deep convolutional neural network (FeaOri-DCNN) for PET image denoising that uses weight maps in order to steer the training toward contrast preservation for small features. To obtain the weight maps, we first manually segment the lesions in the target images to create lesion masks. Lesion voxels are assigned stronger weights than the background voxels followed by a Gaussian smoothing. This weight map is then incorporated into the loss function optimization. We first trained the proposed FeaOri-DCNN and a conventional DCNN built on a five-layer residual network architecture with simulated and phantom images containing hot spheres with various size and contrast. We then trained an eight-layer network with 8 patient studies and 1 phantom study. We evaluated the five-layer network on phantom studies and the eight-layer network on 2 patient studies inserted with GATE simulated lesions. The results of the phantom studies show that FeaOri-DCNN improved contrast recovery on small and low contrast spheres by up to 36% while performing similarly in terms of noise reduction in the background. Similar results were also observed in the patient studies.