Instance-Specific Denoising: A Hybrid Acquisition Framework for Deep Learning in Low-Dose Computed Tomography

Abstract

We propose an innovative approach to low-dose computed tomography that combines a hybrid acquisition protocol with a deep convolutional neural network. Our protocol interleaves sparse, uniformly spaced high-exposure projections with dense low-exposure projections. This strategy provides object-specific training data in situ and ensures the denoising task is always a constrained interpolation between high-quality samples, which robustly prevents model-induced artifacts. The CNN then processes spatio-angular sequences of patches from these projections, explicitly leveraging angular correlation to denoise the radiographs prior to reconstruction. We validated our framework on three new, publicly available CBCT datasets. The results demonstrate a remarkable improvement in image quality at both the radiographic and volumetric levels. Even with overall exposure savings exceeding 90%, our denoised results achieve a very high similarity to the full-exposure ground truth, successfully recovering fine structural details lost in the noisy baseline data. The accompanying datasets provide a valuable resource for future research and benchmarking.