Patient-specific noise power spectrum measurement via generative adversarial networks (Conference Presentation)

A deep learning Generative Adversarial Networks (GANs) were developed and validated to provide an accurate way of direct NPS estimation from a single patient CT scan. GANs were utilized to map a white noise input to a CT noise realization with correct CT noise correlations specific to a single local uniform ROI. To achieve this, a two-stage strategy was developed. In the pre-training stage, ensembles of 64x64 MBIR noise-only images of a quality assurance phantom were used as training samples to jointly train the generator and discriminator. They were fined-tuned using training samples from a single 101x101 ROI of an abdominal anthropomorphic phantom. Results from GANs and physical scans were compared in terms of its mean frequency and radial averaged NPS. This workflow was extended to a patient case where reference dose and 25% of reference dose CT scans were provided for fine-tuning. GANs generated noise-only image samples that are indistinguishable from physical measurement. The overall mean frequency discrepancy between NPS generated from GANs and those from physically acquired data was 0.2% for anthropomorphic phantom validation. The KL divergence for 1D radial averaged NPS profile of these two NPS acquisitions was 2.2×10^(-3). Statistical test indicates trained GANs generated equivalent NPS to physical scans. In clinical patient-specific NPS studies, it showed a distinction between the reference dose case and 25% of reference dose case. It was demonstrated the GANs characterized the properties of CT noise in terms of its mean frequency and 1D NPS profile shape.