Purpose: Low-dose CT (LDCT) imaging is desirable in many clinical applications to reduce X-ray radiation dose to patients. Inspired by deep learning (DL), a recent promising direction of model-based iterative reconstruction (MBIR) methods for LDCT is via optimization-unrolling DL-regularized image reconstruction, where pre-defined image prior is replaced by learnable data-adaptive prior. However, LDCT is clinically multilevel, since clinical scans have different noise levels that depend of scanning site, patient size, and clinical task. Therefore, this work aims to develop an adaptive-hyper-parameter DL-based image reconstruction method (AHP-Net) that can handle multilevel LDCT of different noise levels.Method: AHP-Net unrolls a half-quadratic splitting scheme with learnable image prior built on framelet filter bank, and learns a network that automatically adjusts the hyperparameters for various noise levels. Each stage of the AHP-Net contains one inversion block and a denoising block, where the denoising block is built-on a CNN. The main difference of the proposed AHP-Net from other deep learning solutions lies the design of the inversion block. In the proposed inversion block, we replaced the often-used gradient operator ∇ by the filter banks with 8 high-pass filters from linear spline framelet transform, motivated by its success in ℓ 1 -norm relating regularization in image recovery. Moreover, we proposed to pay special attention to the hyper-parameters involved in the inversion block, and presented a MLP-based NN to predict hyper-parameters that adaptive to both dose level and image content.Result: AHP-Net provides a single universal training model that can handle multilevel LDCT. Extensive experimental evaluations using clinical scans suggest that AHP-Net outperformed conventional MBIR techniques and state-of-the-art deep-learning-based methods for multilevel LDCT of different noise levels.
Conclusions:The experiments showed the advantage of the proposed method over classic non-learning methods and some representative deep learning based methods for LDCT reconstruction. Also, another advantage is that the proposed method can only train a single model with competitive performance to process measurement data with varying dose levels.