Search

Low-dose computed tomography (LDCT) imaging is a valuable tool in medical diagnostics due to its reduced radiation exposure compared to traditional CT scans. However, LDCT images are often plagued by high levels of noise, compromising their diagnostic quality. In this paper, a novel deep learning-based approach is proposed for LDCT image denoising through deep unsupervised domain adaptation (UDA) with self-regularization. The objective is to address the challenges of LDCT image denoising and enhance LDCT image quality in the target domain. For data preprocessing, the LDCT images are selected from the LDCT and projection dataset, LUNA16 dataset, CT low-dose reconstruction dataset and Mayo dataset. The data augmentation and normalization processes come under the data preprocessing to ensure consistency and suitability for the denoising task. The self-regularization techniques such as consistency regularization, pseudo-labeling and entropy minimization are used to contribute model's ability for learning domain-invariant features. The deep unsupervised domain adaptation model effectively bridges the domain gap between the source and target domain. The joint optimization loss function is implemented to ensure that the denoising network extracts and expresses the similarities of patches from the target domain. The local and global information are integrated based on the training process for enhancing the denoising capabilities of the network. The comprehensive analyses are conducted based on significant performance metrics such as peak signal-to-noise ratio (PSNR), structural similarity index, mean absolute error, structural content, feature similarity and mean structural similarity index. The proposed model attained a PSNR rate of 46.21 dB which showed a better performance compared to other existing methods. [ABSTRACT FROM AUTHOR]