Enhancing low-dose CT denoising via multi-view knowledge transfer without paired data

Despite remarkable advancements in deep learning for low-dose computed tomography (LDCT) denoising, two significant challenges persist: (i) the requirement for paired LDCT and high-dose computed tomography (HDCT) images, which are often impractical to obtain in clinical settings; and (ii) the tendency of existing methods to train models using a single axial view, thereby overlooking complementary information from other views and consequently limiting their performance.

To address these issues, we propose a Multi-view-to-Single Knowledge Transfer (MvSKT) framework for unpaired LDCT denoising. Our approach involves splitting the 3D unpaired computed tomography (CT) data into 2D images from various views, including axial, sagittal, and coronal. This allows us to train three view-independent 2D GAN models in an unsupervised manner. By stacking successive 2D outputs from each view-independent model into a volumetric format and splitting them into axial-view images, we generate multiple complementary predictions for each axial CT image. Leveraging these predictions as priors, we transfer multi-view knowledge to a single-view model through pseudo-supervision. This process involves fusing multiple view-complementary predictions into reliable pseudo-images using a cycle-consistency-weighted method.

Extensive experiments on the AAPM-Mayo dataset demonstrate that MvSKT outperforms other unpaired denoising methods and even achieves performance comparable to supervised approaches.

Consequently, the MvSKT framework effectively harnesses multi-view information from unpaired data to enhance LDCT denoising without the strict requirement of paired clinical data.