Unsupervised Deep-learning Methods for Low-dose Computed Tomography Reconstruction

Abstract

Computed tomography (CT) has become an indispensable imaging technique in medical diagnostics and industrial applications, owing to its non-invasive nature and high resolution in visualizing object internal structures. While X-ray CT (X-ray computed tomography) significantly enhances lesion detection capabilities, excessive exposure to X-ray radiation raises substantial health concerns, including elevated cancer risks and potential genetic damage. Although low-dose CT (LDCT) protocols reduce radiation safety concerns, they inevitably introduce severe noise and artifacts, which might compromise diagnostic accuracy. Recent advances in supervised deep learning approaches have demonstrated remarkable success in LDCT reconstruction. However, the reliance on paired training data severely limits their deployment in practical CT applications. This fundamental constraint highlights the critical importance of developing unsupervised reconstruction methods. Although existing unsupervised LDCT methods have made notable progress, they still face challenges requiring systematic solutions. This thesis makes three contributions to advance the field of unsupervised LDCT reconstruction: 1. Current dual-domain self-supervised LDCT denoising methods typically neglect the heterogeneity of the non-stationary Gaussian noise levels in low-dose sinograms and usually treat them as common images without appropriately controlling the denoising strength. The denoisers employed by them, which are based on classic convolutional neural networks (CNN), will lead to blurring artifacts in the reconstructed images if directly used for sinogram denoising. In addition, the denoising strength in the sinogram and image domains must be well-balanced to avoid introducing over-blurring or secondary artifacts in the reconstructed images, but existing approaches do not focus on this crucial point. To address these limitations, this thesis proposes a novel end-to-end dual-domain self-supervised framework for LDCT denoising. It employs Dropblock layers to adaptively localize the effect of convolution for sinogram denoising and sets a weighted average between the denoised sinograms and their noisy counterparts to better control the denoising strength, thus effectively reducing the blurring artifacts and leading to a well-balanced dual-domain denoising. Numerical experiments demonstrate the effectiveness and superior performance of the proposed method. 2. Existing normalizing flows (NFs)-based unsupervised LDCT reconstruction methods face challenges in both image quality and computational efficiency. On one hand, they typically utilize a two-way transformation strategy between noisy images and latent variables, which could easily lead to detail loss and secondary artifacts in the reconstructed images. On the other hand, training NFs on high-resolution CT images is computationally intensive. Although conditional normalizing flows (CNFs) can mitigate computational costs by learning conditional probabilities, existing approaches rely on labeled data for conditionalization, leaving unsupervised CNFs-based LDCT reconstruction a challenge. To tackle these issues, this thesis proposes a novel unsupervised LDCT iterative reconstruction algorithm based on CNFs. The proposed method employs a strict one-way transformation strategy during the alternating optimization in the dual spaces to prevent detail loss and secondary artifacts, and proposes a novel unsupervised conditionalization strategy for CNFs, thus achieving efficient training and inference on high-resolution images. The proposed method illustrates high-quality and relatively fast unsupervised reconstruction. Experiments across two datasets demonstrate the superior performance of the proposed method by rivaling several state-of-the-art methods. 3. Current diffusion model-based LDCT reconstruction methods solely employ the prior distributions learned from normal-dose CT data but neglect the valuable priors in low-dose data, which contain information about the characteristics of low-dose noise and artifacts. In addition, their reconstruction speed suffers from multiple sampling steps, presenting low efficiency. Breaking new ground in prior utilization, this thesis proposes a novel unsupervised LDCT iterative reconstruction algorithm based on dual denoising diffusion probabilistic models (DDPM), which leverages both normal-dose and low-dose priors. Specifically, the proposed method employs two DDPMs to learn the prior distributions from normal-dose and low-dose images, respectively, and incorporates them into a joint iterative reconstruction framework. To accelerate the reconstruction, partial diffusion sampling and single-pass deterministic reconstruction strategy are utilized in the proposed method, as well as the spaced OS-SART approaches. Experiments at different dose levels demonstrate the outstanding performance and fast speed of the proposed reconstruction algorithm. Through these innovations and contributions, this thesis addresses some significant challenges in current deep-learning-based unsupervised LDCT reconstruction methods, and proposes three novel unsupervised reconstruction algorithms. The proposed methods effectively improve the performance and efficiency of LDCT reconstruction under different unsupervised learning conditions, and provide some new perspectives and inspirations for future research.

Ähnliche Arbeiten