Abstract
The state-of-the-art x-ray photon-counting CT (PCCT) generates images in multi-energy bins simultaneously
with high spatial resolution and low radiation dose for tissue characterization and material decomposition. FDA
has approved the techniques in 2021. Both clinical PCCT and micro-PCCT scanners are now commercially
available. This opens a new door to opportunities for functional, cellular, and molecular x-ray imaging with novel
contrast agents such as bismuth and gold nanoparticles. However, x-ray photon-counting detectors are not
perfect, and it remains challenging to reconstruct high-quality PCCT images for various clinical applications.
Over the past several years, deep learning-based tomographic imaging has become a new frontier of image
reconstruction. Different from compressive sensing (CS) methods, which totally rely on the prior information in
terms of an accurate mathematical constraint, the emerging deep learning-based approach is empowered by big
data with which a deep network can be trained for superior tomographic reconstruction. However, a recent study
published in PNAS revealed three types of instabilities of deep tomographic reconstruction networks, which are
believed to be fundamental due to lack of kernel awareness and “nontrivial to overcome”, but CS-based
reconstruction was reported in that study to be stable because of its kernel awareness. Meanwhile, it is hard to
collect large amounts of data with ground-truths for supervised network training up to the clinical image quality.
To overcome the aforementioned challenges in the context of a clinical trial with PCCT using Medipix detectors,
our overall goal is to develop an Unsupervised Deep Learning Approach (UDLA) for few-view and low-dose
image reconstruction based on our Analytic Compressive Iterative Deep (ACID) architecture but specific to PCCT
data, with much higher spatial resolution and computational efficiency, and without the requirement of ground-
truth for training. ACID combines the data-driven power of deep learning, the kernel-awareness of CS, and
iterative refinement to deliver image reconstruction results accurately and stably. To achieve our goal, three
specific aims are defined as follows. Aim 1: UDLA will be designed, developed, optimized, and integrated into
an open-source platform, including a deep end-to-end reconstruction network and an advanced CS module with
a multi-constraint model; Aim 2: UDLA will be tested for stability and generalizability, and accelerated via
software optimization on a high-performance computing platform; and Aim 3: UDLA will be evaluated and
validated in simulation, experiments, and retrospective use of clinical extremity imaging PCCT data.
Upon the completion of this project, the UDLA software should have been characterized for clinical extremity
imaging using Medpix-based PCCT to outperform contemporary iterative algorithms, without the vulnerabilities
of existing deep reconstruction networks and the requirements of ground-truth for network training. In a broader
perspective, our approach represents a paradigm shift towards the integration of model-based and data-driven
reconstruction methods, and may have a lasting impact on PCCT and other tomographic imaging modalities.
