Zheng Chang

Overview:

Dr. Chang's research interests include radiation therapy treatment assessment using MR quantitative imaging, image guided radiation therapy (IGRT), fast MR imaging using parallel imaging and strategic phase encoding, and motion management for IGRT.

Positions:

Professor of Radiation Oncology

Radiation Oncology
School of Medicine

Member of the Duke Cancer Institute

Duke Cancer Institute
School of Medicine

Education:

Ph.D. 2006

University of British Columbia

Grants:

Publications:

An investigation of machine learning methods in delta-radiomics feature analysis.

PURPOSE: This study aimed to investigate the effectiveness of using delta-radiomics to predict overall survival (OS) for patients with recurrent malignant gliomas treated by concurrent stereotactic radiosurgery and bevacizumab, and to investigate the effectiveness of machine learning methods for delta-radiomics feature selection and building classification models. METHODS: The pre-treatment, one-week post-treatment, and two-month post-treatment T1 and T2 fluid-attenuated inversion recovery (FLAIR) MRI were acquired. 61 radiomic features (intensity histogram-based, morphological, and texture features) were extracted from the gross tumor volume in each image. Delta-radiomics were calculated between the pre-treatment and post-treatment features. Univariate Cox regression and 3 multivariate machine learning methods (L1-regularized logistic regression [L1-LR], random forest [RF] or neural networks [NN]) were used to select a reduced number of features, and 7 machine learning methods (L1-LR, L2-LR, RF, NN, kernel support vector machine [KSVM], linear support vector machine [LSVM], or naïve bayes [NB]) was used to build classification models for predicting OS. The performances of the total 21 model combinations built based on single-time-point radiomics (pre-treatment, one-week post-treatment, and two-month post-treatment) and delta-radiomics were evaluated by the area under the receiver operating characteristic curve (AUC). RESULTS: For a small cohort of 12 patients, delta-radiomics resulted in significantly higher AUC than pre-treatment radiomics (p-value<0.01). One-week/two-month delta-features resulted in significantly higher AUC (p-value<0.01) than the one-week/two-month post-treatment features, respectively. 18/21 model combinations were with higher AUC from one-week delta-features than two-month delta-features. With one-week delta-features, RF feature selector + KSVM classifier and RF feature selector + NN classifier showed the highest AUC of 0.889. CONCLUSIONS: The results indicated that delta-features could potentially provide better treatment assessment than single-time-point features. The treatment assessment is substantially affected by the time point for computing the delta-features and the combination of machine learning methods for feature selection and classification.
Authors
MLA Citation
Chang, Yushi, et al. “An investigation of machine learning methods in delta-radiomics feature analysis.Plos One, vol. 14, no. 12, 2019, p. e0226348. Pubmed, doi:10.1371/journal.pone.0226348.
URI
https://scholars.duke.edu/individual/pub1423241
PMID
31834910
Source
pubmed
Published In
Plos One
Volume
14
Published Date
Start Page
e0226348
DOI
10.1371/journal.pone.0226348

WE-C-116-07: Tumor Enhancement Using Deformable Image Registration for Four-Dimensional Magnetic Resonance Imaging (4D-MRI): A Feasibility Study.

PURPOSE: We have previously developed a 4D-MRI technique using the fast imaging sequence employing steady-state acquisition (FIESTA) sequence, which has suboptimal tumor-to-tissue contrast-to-noise ratio (CNR) due to its T2*/T1 weighting. This study investigated the feasibility of enhancing the tumor-to-tissue CNR using deformable image registration (DIR). METHODS: Five patients with cancers in the liver were included in an IRB-approved study. 4D-MRI images were acquired on a 1.5T GE scanner and reconstructed off line using in-house developed program. All patients were also imaged with a T2-w fast recovery fast spin-echo (FRFSE) sequence at the end-of-exhalation phase. Deformation vectors between respiratory phases of the 4D-MRI were determined using commercial software. Pseudo 'enhanced' 4D-MRI was then generated by applying the deformation vectors to the T2-w FRFSE MR images. Motion trajectories of tumor and diaphragm and tumor-to-tissue CNR were compared between the original T2*/T1-w 4D-MRI and the 'enhanced' T2-w 4D-MRI. To validate our method, we performed a simulation study based on a 4D digital human phantom. MR images with T2*/T1-w and T2-w with were generated by assigning organ intensities corresponding to those in FIESTA and FRFSE images, respectively. RESULTS: In the phantom study, motion trajectories of the hypothesized 'tumor' matched excellently between the original T2*/T1-w 4D-MRI and the 'enhanced' T2-w 4D-MRI. Mean(±SD) absolute difference in motion amplitude was 0.66 (±0.62) mm. In the patient study, tumor and diaphragm motion trajectories closely matched between the two 4D-MRIs: mean correlation coefficient was great than 0.97 in all directions; the mean (±SD) absolute difference in motion amplitude was smaller than 0.55(±0.19) mm. Tumor-to-tissue CNR was significantly improved from 7.57(±5.6) in the original 4D-MRI to 23.75(±15.8) in the 'enhanced' 4D-MRI. CONCLUSION: It is feasible to improve tumor-to-tissue CNR of T2*/T1-w 4D-MRI using the DIR method. The 'enhanced' 4D-MRI retained comparable tumor motion information as the original 4D-MRI. This work is partly supported by funding from NIH (1R21CA165384-01A1) and a research grant from the Golfers Against Cancer (GAC) Foundation.
Authors
MLA Citation
Yang, J., et al. “WE-C-116-07: Tumor Enhancement Using Deformable Image Registration for Four-Dimensional Magnetic Resonance Imaging (4D-MRI): A Feasibility Study.Med Phys, vol. 40, no. 6Part29, June 2013, p. 485. Pubmed, doi:10.1118/1.4815569.
URI
https://scholars.duke.edu/individual/pub1164329
PMID
28518641
Source
pubmed
Published In
Med Phys
Volume
40
Published Date
Start Page
485
DOI
10.1118/1.4815569

Assessing Effects of Ion Collection Efficiency in Flattening Filter-Free (FFF) Beams On Three TrueBeam Machines

Authors
Chang, Z; Wu, Q; Adamson, J; Ren, L; Bowsher, J; Yan, H; Thomas, A; Yin, F
MLA Citation
Chang, Z., et al. “Assessing Effects of Ion Collection Efficiency in Flattening Filter-Free (FFF) Beams On Three TrueBeam Machines.” Medical Physics, vol. 39, no. 6, 2012, pp. 3719–3719. Wos-lite, doi:10.1118/1.4735130.
URI
https://scholars.duke.edu/individual/pub1266162
Source
wos-lite
Published In
Medical Physics
Volume
39
Published Date
Start Page
3719
End Page
3719
DOI
10.1118/1.4735130

Evaluation of integrated respiratory gating systems on a Novalis Tx system

The purpose of this study was to investigate the accuracy of motion tracking and radiation delivery control of integrated gating systems on a Novalis Tx system. The study was performed on a Novalis Tx system, which is equipped with Varian Real-time Position Management (RPM) system, and BrainLAB ExacTrac gating systems. In this study, the two systems were assessed on accuracy of both motion tracking and radiation delivery control. To evaluate motion tracking, two artificial motion profiles and five patients' respiratory profiles were used. The motion trajectories acquired by the two gating systems were compared against the references. To assess radiation delivery control, time delays were measured using a single-exposure method. More specifically, radiation is delivered with a 4 mm diameter cone within the phase range of 10%-45% for the BrainLAB ExacTrac system, and within the phase range of 0%-25% for the Varian RPM system during expiration, each for three times. Radiochromic films were used to record the radiation exposures and to calculate the time delays. In the work, the discrepancies were quantified using the parameters of mean and standard deviation (SD). Pearson's product-moment correlational analysis was used to test correlation of the data, which is quantified using a parameter of r. The trajectory profiles acquired by the gating systems show good agreement with those reference profiles. A quantitative analysis shows that the average mean discrepancies between BrainLAB ExacTrac system and known references are 1.5 mm and 1.9 mm for artificial and patient profiles, with the maximum motion amplitude of 28.0 mm. As for the Varian RPM system, the corresponding average mean discrepancies are 1.1 mm and 1.7 mm for artificial and patient profiles. With the proposed single-exposure method, the time delays are found to be 0.20 ± 0.03 seconds and 0.09 ± 0.01 seconds for BrainLAB ExacTrac and Varian RPM systems, respectively. The results indicate the systems can track motion and control radiation delivery with reasonable accuracy. The proposed single-exposure method has been demonstrated to be feasible in measuring time delay efficiently.
Authors
Chang, Z; Liu, TH; Cai, J; Chen, Q; Wang, Z; Yin, FF
MLA Citation
Chang, Z., et al. “Evaluation of integrated respiratory gating systems on a Novalis Tx system.” Journal of Applied Clinical Medical Physics, vol. 12, no. 3, Jan. 2011, pp. 71–79. Scopus, doi:10.1120/jacmp.v12i3.3495.
URI
https://scholars.duke.edu/individual/pub763048
Source
scopus
Published In
Journal of Applied Clinical Medical Physics
Volume
12
Published Date
Start Page
71
End Page
79
DOI
10.1120/jacmp.v12i3.3495

SU‐GG‐I‐24: Improving IGRT Efficiency Using GPU‐Based Ultrafast Reconstruction of DTS/CBCT and DRR

Purpose: One challenge of on‐line image‐guided radiation therapy is requirement of immediate availability of reconstructed 3D or 4D images for evaluation. This study is to develop a highly data‐parallel version and platform of accelerating reconstruction for digital tomosynthesis (DTS)/Cone‐Beam CT (CBCT) and digitally‐reconstructed radiographs (DRR) based on ray‐tracing algorithm. Method and Materials: A modified FDK algorithm of tomographic image reconstruction is being implemented on a CUDA‐based commercial graphics processing unit (GPU). This algorithm is developed to address time‐consuming process of reconstructing DTS/ CBCT, and DRR. The original projection images are first allocated to the textures on the graphic card. Based on the cone‐beam geometry, the projection matrix for each gantry rotation is generated from the ray intersection and overlaid to the GPU memory. A unique thread ID will be assigned to each voxel volume for further computation. Global computation kernel of FDK processing is called from shared memory of each block on GPU. The reconstructed 3D data is then transferred back to CPU after synchronization. Image quality of the reconstructed images with both hardware and software techniques is compared using differential contrast‐to‐singal ratio (CNR) and pixel signal‐to‐noise ratio (PSNR) for various clinical sites including lung, pelvis and head&neck. Results: With the new algorithm, the times for reconstructing DTS and DRR are improved by a factor of 100 using Quadro FX5800 and 30 using GT9600, respectively, over the conventional software method. The DTS/CBCT for a lung patient can be reconstructed and rendered with 4 to 51 seconds dependent on the number of projections. Consistency comparison of images reconstructed using both hardware and software techniques shows greater than both consistancy and image resolution increasing. Conclusion: A GPU‐accelerated ultra‐fast reconstruction allows real‐time on‐board imaging process to be completed within a minute to improve the efficiency of IGRT applications. © 2010, American Association of Physicists in Medicine. All rights reserved.
Authors
Jian, Y; Godfrey, D; Chang, Z; Yin, F
MLA Citation
Jian, Y., et al. “SU‐GG‐I‐24: Improving IGRT Efficiency Using GPU‐Based Ultrafast Reconstruction of DTS/CBCT and DRR.” Medical Physics, vol. 37, no. 6, 2010, p. 3106. Scopus, doi:10.1118/1.3468057.
URI
https://scholars.duke.edu/individual/pub1266378
Source
scopus
Published In
Medical Physics
Volume
37
Published Date
Start Page
3106
DOI
10.1118/1.3468057