In positron emission tomography (PET), decreasing patient injected radiation dose or scan time leads to a reduction in acquisition counts and noisy reconstructed images. Approaches such as post-smoothing (PS) and total variation (TV) denoising can help reduce noise at the cost of resolution or bias, respectively. However, recent advances in convolutional neural networks (CNNs) have lead to promising results in low count PET. Some of these can take advantage of jointly acquired data from other modalities such as magnetic resonance (MR) imaging. However, there are many considerations when designing an architecture depending on the specific task and data available. This work compares some of the current state-of-the-art approaches. Results demonstrate that more powerful, complex networks do not necessarily yield better performance. Normalised root mean squared error (NRMSE) is decreased by 85.5% compared to standard maximum likelihood expectation maximisation (MLEM) PET reconstruction by using a 3-layer micro-network on very low (3.01M) count simulation data. This compares to reductions of 80.2% and 85.3% when total variation (TV) edge enhancement and an adversarial discriminator are incorporated into the loss function, respectively. Using both TV and a discriminator term produces a slightly better whole-brain NRMSE reduction of 86.0%, though with a risk of false positive lesions. This indicates that more investigation is required to determine general rules for choosing architectures and hyperparameters for low count PET.
Paper
Full text
3D Convolutional Adversarial Micro-networks for Low Count PET-MR Post-processing
Semantic Scholar · Medicine · 2019
Abstract
In positron emission tomography (PET), decreasing patient injected radiation dose or scan time leads to a reduction in acquisition counts and noisy reconstructed images. Approaches such as post-smoothing (PS) and total variation (TV) denoising can help reduce noise at the cost of resolution or bias, respectively. However, recent advances in convolutional neural networks (CNNs) have lead to promising results in low count PET. Some of these can take advantage of jointly acquired data from other modalities such as magnetic resonance (MR) imaging. However, there are many considerations when designing an architecture depending on the specific task and data available. This work compares some of the current state-of-the-art approaches. Results demonstrate that more powerful, complex networks do not necessarily yield better performance. Normalised root mean squared error (NRMSE) is decreased by 85.5% compared to standard maximum likelihood expectation maximisation (MLEM) PET reconstruction by using a 3-layer micro-network on very low (3.01M) count simulation data. This compares to reductions of 80.2% and 85.3% when total variation (TV) edge enhancement and an adversarial discriminator are incorporated into the loss function, respectively. Using both TV and a discriminator term produces a slightly better whole-brain NRMSE reduction of 86.0%, though with a risk of false positive lesions. This indicates that more investigation is required to determine general rules for choosing architectures and hyperparameters for low count PET.