|
Spectral Computed Tomography (CT) is a versatile imaging technique increasingly utilized in industry, medicine, and scientific research. This technique allows us to observe the energy-dependent X-ray attenuation throughout an object by using Photon Counting Detector (PCD) technology. However, a major drawback of Spectral CT is the increase in noise due to a lower photon count per channel, as increasing the number of energy channels without also increasing scan time reduces the photon count per channel. This challenge often complicates quantitative material identification, which is a major application of the technology. In this study, we investigate the use of unsupervised image denoising approaches and demonstrate the applicability of the Noise2Inverse method, an unsupervised denoising method for tomographic imaging. These approaches have the advantage over supervised machine learning methods in that they do not require any additional clean or noisy training data, which can be very difficult to collect in Spectral CT imaging. Our model uses a U-Net paired with a block-based training approach. In particular, we demonstrate that the block-based models can be efficiently trained using small image blocks, each block incorporating spectral information. This training process is performed on images that are reconstructed from subsets of measured Spectral tomography data. The experiments used two simulated Spectral CT phantoms, each with a unique shape and material decomposition. Upon evaluation using the peak signal-to-noise ratio (PSNR) and structural similarity index (SSIM) performance metrics, our approach exhibited improvements compared to two alternative approaches: the unsupervised Low2High method previously employed in sparse Spectral CT imaging and a traditional Iterative reconstruction method that imposes a Total Variation (TV) constraint.
|
