Chengzhu Zhang, Ph.D.

Chengzhu Zhang, Ph.D.

He/His

Medical Physics Resident

Selected Publication

Overview

Below you can find my comments and thoughts for each of the research project I worked on during my Ph.D. For simplified overview, You can directly go to my Google Scholar profile.

First-Author Papers

(Editor’s Choice) Noise power spectrum (NPS) in computed tomography: Enabling local NPS measurement without stationarity and ergodicity assumptions

Published:

This paper serves as the foundational paper for accurate and robust noise power spectrum (NPS) measurement. It challenged the dogma of conventional NPS measurement by proposing a completely differnt pathway. This paper pointed out that the conventional way can be decomposed into diagonal and off-diagonal contributions. It is the diagonal component that contributes to the meaningful signal while the off-diagonal component can contribute to purely noise if the detector correlation can be measured to be low. This paper received editor’s choice in the published issue.

Recommended citation: Zhang, Chengzhu, et al. "Noise power spectrum (NPS) in computed tomography: Enabling local NPS measurement without stationarity and ergodicity assumptions." Medical physics (2024).
Download Paper

Keywords: Local Noise Power Spectrum, Experiment Measurement

Experimental measurement of local noise power spectrum (NPS) in photon counting detector‐CT (PCD‐CT) using a single data acquisition

Published:

This paper opens the door to patient-specific CT image quality assessment by harnessing the power of PCD-CT data acquisition. Traditionally, multiple repeated scans on a surrogate phantom has to be conducted to rigorously measure the noise power spectrum (NPS). This is strictly prohibited on the patient. However, with the merging PCD-CT technology, patient-specific noise power spectrum can be measured with single CT data acquisition under the proposed new paradigm. Its proceeding paper received Robert F. Wagner All-Conference Best Student Paper Award (Runner-up) at SPIE 2023.

Recommended citation: Zhang, Chengzhu, et al. "Experimental measurement of local noise power spectrum (NPS) in photon counting detector‐CT (PCD‐CT) using a single data acquisition." Medical physics (2024).
Download Paper

Keywords: Patient-specific image quality assessment, Photon-counting CT, Classic mathematics for CT image reconstruction

Deep-Interior: A new pathway to interior tomographic image reconstruction via a weighted backprojection and deep learning

Published:

This paper is one of the highlights in my PhD career. Interior problem is a notoriously difficult problem in the history of CT reconstruction. Mathematicians and medical physicists have established rigorous theoretical frameworks to solve the interior problem. Among them, the fully truncated CT data reconstruction is a whole another level. With rapid devleopment of AI enabled reconstruction, we intensly derived a unique feature space called the backprojection space for the deep neural network to learn the necessary deconvolution scheme to reconstruction ROI as small as 5 cm in diameter regardless of its position.

Recommended citation: Zhang, Chengzhu, and Guang‐Hong Chen. "Deep‐Interior: A new pathway to interior tomographic image reconstruction via a weighted backprojection and deep learning." Medical physics 51.2 (2024): 946-963.
Download Paper

Keywords: Interior Problem, Backprojection filtering for divergent beam geometry

Reconstruction of three‐dimensional tomographic patient models for radiation dose modulation in CT from two scout views using deep learning

Published:

ScoutCT-Net CT is a very cool deep learning model my colleage Juan and I developed. It was intended for reconstruction of 3D volumetric image with soft-tissue contrast from only two orthogonal views. It is the pioneer work in this category that even predated the famous single-view paper. It is a powerful tool and can be applied to both diagnostic and radiation therapy settings. In ths published paper, we demonstrated its powerful generative potential to convert orthorgonal scout views to 3D CT. The results were quantitatively analyzed in the clinical endpoint of fluence modulation and dose verification.

Recommended citation: Montoya, Juan C., et al. "Reconstruction of three‐dimensional tomographic patient models for radiation dose modulation in CT from two scout views using deep learning." Medical physics 49.2 (2022): 901-916.
Download Paper

Keywords: Two-view volumetric reconstruction, Fluence modulation

Accurate and robust sparse‐view angle CT image reconstruction using deep learning and prior image constrained compressed sensing (DL‐PICCS)

Published:

DL-PICCS is my first peer-reviewed work in my PhD career. At that time, there was an arms race in the field of deep learning-based reconstruction. The purpose of this works to warn the community the contridiction of personalized nature of CT image reconstruction and the statistical regression nature of the deep learning model. DL-PICCS was born to enforce the deep learning models to be data consistent without having to build the data consistency module into the network. It utilized the framework of the well-known PICCS method developed by my PhD advisor. The deep leanring reconstructed image is served as an prior image and the final image is jointly optimized by both the deep learning prior image and a handcrafted regulalizer from the wisdom of compressed sensing in the past 15 years. In this paper, DL-PICCS corrects false positive and false negative anatomical structures and significantly boosts the image quality.

Recommended citation: Zhang, Chengzhu, Yinsheng Li, and Guang‐Hong Chen. "Accurate and robust sparse‐view angle CT image reconstruction using deep learning and prior image constrained compressed sensing (DL‐PICCS)." Medical physics 48.10 (2021): 5765-5781.
Download Paper

Keywords: Generalizability and robustness of AI imaging, Prior knowledge constrained deep learning

CT artifact reduction via U-net CNN

Published:

This paper is a pioneer work from my undergraduate research in the field of deep learning CT image reconstruction. It address the challenging problem of metal artifact in CT which even now still bothers both the diagnostic and therapeutic physicists in the conventional setting. In this paper, clinical data from dental CT with metla implants were used to validate our proposed model.

Recommended citation: C. Zhang and Y. Xing "CT artifact reduction via U-net CNN", Proc. SPIE 10574, Medical Imaging 2018: Image Processing, 105741R (5 March 2018); https://doi.org/10.1117/12.2293903
Download Paper

Keywords: AI-based image reconstruction, Metal artifact reduction

Highlighted Co-author Papers

High pitch helical CT reconstruction

Published:

High-pitch helical CT is a difficult problem consisting of two types of imaging tasks: limited-view reconstruction and sparse-view reconstruction. The reconstruction framework of DL-PICCS developed in my first-authored paper was used to tackle this problem.

Recommended citation: Hayes, John W., et al. "High pitch helical CT reconstruction." IEEE Transactions on Medical Imaging 40.11 (2021): 3077-3088.
Download Paper

Keywords: High-pitch helical CT, Prior knowledge constrained deep learning

Learning to reconstruct computed tomography images directly from sinogram data under a variety of data acquisition conditions

Published:

iCT-Net is a such a cool network that was inspired by the filtered-backprojection algorithm but with everything learned through data. It is an end-to-end reconstruction network that consists of three major modules, data rectification module, feature learning module and backprojection module. The data rectification module preprocesses the noisy data from a real CT scanner. The feature learning module learns the essential filtering process and beyond to avoid data contamination in the traditional convolutional process. Then backprojection module enables domain transform and end-to-end learning. The proposed iCT-Net is an intelligent reconstruction framework to address multiple challenging reconstruction problems.

Recommended citation: Li, Yinsheng, et al. "Learning to reconstruct computed tomography images directly from sinogram data under a variety of data acquisition conditions." IEEE transactions on medical imaging 38.10 (2019): 2469-2481.
Download Paper

Keywords: End-to-end AI-driven image reconstruction, Interior problem

Book Chapters