Self-Supervised Noise-Aware Kernel Synthesis for Improved X-ray Computed Tomography Imaging

• Published in: 2024 IEEE International Symposium on Biomedical Imaging (ISBI) pp. 1 - 4
• Main Authors: Aggarwal, Hemant Kumar, Yalavarthy, Phaneendra K., Langoju, Rajesh
• Format: Conference Proceeding
• Language: English
• Published: IEEE 27.05.2024
• Subjects: and Self-supervised Learning; Computed tomography; Computer architecture; Image quality; Image Reconstruction; Image Synthesis; Inverse problems; Kernel; Lung; Noise
• ISSN: 1945-8452
• DOI: 10.1109/ISBI56570.2024.10635296

The process of image-based kernel synthesis (KS) transforms X-ray computed tomography (CT) images reconstructed with one type of kernel into images of another type of kernel without requiring the original sinogram data. The kernel synthesis approach enhances computer aided detection, low contrast distinguishability, and quantitative analysis. However, this process can also intensify the noise in input images, leading to a degradation in image quality. Noise-aware kernel synthesis is a challenging task that requires prior knowledge or a regularization function to manage the ill-posedness and noise of the inverse problem. In this study, a self-supervised kernel synthesis (SSKS) method was introduced that explicitly considers the physics of kernel synthesis. It includes information about the modulation transfer function (MTF), resolution, and display fields of view (DFoV). The proposed method utilizes this physics information together with Neumann networks trained with deep image prior for noise aware self-supervised kernel synthesis. In particular, Neumann architecture leads to dense skip connections to improve the sharpness while learning weights of a deep regularizer. Deep image prior (DIP) helps to learn the prior in an self-supervised manner, while controlling the noise using a deep network as implicit prior. The proposed method was tested for lung and bone kernel synthesis from standard and detail kernel images respectively. The results showed that the proposed method could reconstruct an image that has the best features of both input and output kernels, improving the image quality and noise suppression.