Generation of training dataset for deep-learning noise reduction

Abstract

Deep-learning (DL) is used extensively in the noise reduction of x-ray computed tomography. In DL, the availability of ground-truth images (GTI) is critical. GTI contains all anatomical information of the low-dose dataset at a much-reduced noise level, resembling that of a high-dose scan of the object. Clinical GTI, however, is difficult to obtain because of ethical and practical constraints. Although the use of model-based-iterative-reconstruction (MBIR) has been proposed, resulting DL images inherent undesired noise characteristics of MBIR.We propose an approach to utilize low-dose FBP images to generate GTI. The approach is based on the observation that the noise in a thick-slice image (TSI) is suppressed while good texture is maintained. Unfortunately, anatomical details of TSI are degraded, resulting from z-averaging. The proposed approach takes the subtraction-image (SI) between the original and TSI, and performs iterative noise reduction (guided by the standard deviation of SI) to gradually remove noise in SI. After the processing, only the differential anatomical structure between the original and TSI remains. This structure is then subtracted from TSI to arrive at GTI. The amount of noise reduction can be adjusted by modifying the slice-thickness of TSI (a factor of 2.4 noise reduction is selected for this study). This approach is evaluated extensively with both clinical neural and body images. For robustness evaluation, all parameters were kept unchanged. For better visual inspection, difference images between the original and GTI are evaluated and negligible residual structure can be detected.