Fourier analysis of multi-scale neural networks implemented for high-resolution X-ray radiography

Abstract

To realize high-resolution X-ray imaging, this study trains deep neural networks comprising multi-scale encoder/decoder architectures with actual data measured using a low-resolution imaging detector. The network performance is evaluated in the Fourier domain through measurements of the modulation-transfer function (MTF) and noise-power spectrum (NPS), which decompose the contrast and noise power (or variance) in terms of the spatial-frequency components. The designed networks are found to successfully deblur blurry images. The MTF analysis reflects the conventional mean-squared errors and structural similarities within it. Moreover, it indicates the over-regression property of the networks that could not be addressed via the conventional metrics. The NPS analysis reveals that the networks suppressed high-frequency noise. The Fourier analysis shows that the designed networks attempted to match the fidelity between the input and label, while suppressing noise.