Denoising of low-dose chest computed tomography images using a U-net based convolutional Autoencoder and transfer learning
Abstract
OBJECTIVE: Low-Dose Computed Tomography (LDCT) is a widely used imaging modality to perform CT examinations with a reduced radiation exposure to patients, but it is affected by increased image noise and artifacts with respect to standard dose imaging, which can hinder its diagnostic power. Deep learning (DL) approaches have proven effective in LDCT denoising; however, they typically require large training datasets, which are difficult to obtain in clinical settings, particularly for denoising tasks. In this paper, a lightweight and versatile DL-based algorithm for chest LDCT denoising is proposed, trained using a two-step strategy to address the challenge of limited training data. 
Approach. We introduce a U-Net-based Convolutional Autoencoder (UNbCAE) designed to reduce noise in chest LDCT scans while preserving anatomical structures and subtle low-contrast features. The model is first trained on phantom images and then adapted to clinical data using a transfer learning approach, thereby reducing reliance on large clinically relevant datasets populated by low-dose and standard-dose paired images. 
Main results. Quantitative and qualitative experiments conducted on the LUNA16 dataset demonstrate that the proposed method enhances image quality and improves the detectability of pulmonary nodules. Furthermore, our training strategy achieved performance comparable to or exceeding that of current state-of-the-art denoising techniques, yielding an average noise reduction factor of (3.4 +/- 0.6). 
Significance. The UNbCAE model provides high-quality LDCT denoising while mitigating the issues related to the dataset size limitation through an effective transfer learning scheme, minimizing dependence on simulated data and supporting clinical applicability.