Shraddha Pandey1,2, Tugce Kutuk3, Matthew N Mills3, Mahmoud Abdalah4, Olya Stringfield4, Kujtim Latifi3, Timothy J Robinson3, Wilfrido Moreno1, Kamran A Ahmed3, and Natarajan Raghunand2,5
1Electrical Engineering, University of South Florida, Tampa, FL, United States, 2Department of Cancer Physiology, H. Lee Moffitt Cancer Center and Research Institute, Tampa, FL, United States, 3Department of Radiation Oncology, H. Lee Moffitt Cancer Center and Research Institute, Tampa, FL, United States, 4Quantitative Imaging Shared Service, H. Lee Moffitt Cancer Center and Research Institute, Tampa, FL, United States, 5Department of Oncologic Sciences, University of South Florida, Tampa, FL, United States
Synopsis
Stereotactic Radiosurgery (SRS) of asymptomatic brain metastases
provides lasting tumor control with only minor side effects to healthy brain.
An active research area is the development of models to predict tumor response
to a given dose of Radiation Treatment (RT) from analysis of pre-RT and post-RT
MR images (i.e., the forward problem). Here we propose an approach to train a
deep neural net on pre-RT MR images of patients with Breast Cancer Metastases to
the Brain (BCMB), for predicting RT dose maps that will yield desired/target
tumor voxel intensities on post-RT MR images (i.e., the inverse problem).
Purpose
Treatment
planning is a time-consuming and labor-intensive, but critical, process in
radiotherapy of cancer. Some of the manual aspects of treatment planning can be
alleviated with automated solutions that use rule-based, atlas-based and
prior knowledge-based methods to plan the delivery of the prescribed radiation
dose to targeted tumors while sparing organs-at-risk (OAR). However, such
automated treatment planning solutions typically do not provide voxel-level
predictions of RT dose that would be optimal for specific outcomes [1-6].
Voxel-level treatment planning would be beneficial for two tasks: one would be
a reduction in labor-intensive steps to contour tumors and OARs, and the second
would be for prescribing heterogeneous dose distributions to achieve optimal response
within solid tumors that are spatially heterogeneous. Models to accomplish the
first task require training data with ground truth pertaining to normal and
pathologic anatomy, while models to accomplish the second task require ground
truth on voxel-level tumor response to RT. Multiparametric MRI (mpMRI) images, particularly
Apparent Diffusion Coefficient (ADC) of water maps, contain a wealth of
information that is mechanistically relatable to voxel-level tumor response to
therapies [7]. Here we have analyzed standard-of-care T1-weighted unenhanced
(T1w) and contrast-enhanced (T1wCE), T2-weighted (T2w), Fluid-Attenuated
Inversion Recovery (FLAIR) images and ADC maps acquired pre-RT and post-RT in
patients with Breast Cancer Metastases to the Brain
(BCMB) who received SRS. Our first aim is to identify post-RT vs. pre-RT
changes on mpMRI that are correlated with response to SRS in BCMB. Our
second aim is to use these putative response criteria as the voxel-level
outcome measure for training a deep neural network on pre-RT T1w, T1wCE, T2w, FLAIR images and ADC maps to predict the optimal RT
dose map for producing a prescribed/desired voxel-level change within the tumor volume on each
image type.Method
In
this retrospective IRB-approved study, planning CT images and associated RT
dose maps, and T1w, T1wCE, T2w, FLAIR images and ADC maps acquired pre-RT,
post-RT (15-158 days), and at tumor recurrence (54-831 days) in twenty-five BCMB patients
were curated from our Radiology and Radiation Oncology databases. These
patients received a mean SRS dose of 21 Gy (range:15-30 Gy) in 1-5 fractions. mpMRI
images from all scan dates were co-registered to the planning CT using MIRADA-RTx
(Mirada Medical, Denver, CO, USA). Gross Tumor Volume (GTV) contours and the RT
dose map associated with the planning CT could be applied to the mpMRI images
after co-registration. Voxel intensities on T2w, FLAIR, T1w and T1wCE images
were calibrated using two reference normal tissues [8]. Intensity-calibrated
voxels on co-registered mpMRI images were assigned to objectively-defined
tissue types [9]. Pre-RT vs. post-RT changes in voxels within the GTV on ADC
and calibrated T2w, FLAIR, T1w and T1wCE images were analyzed for differences
between responding and recurrent BCMB lesions. We modified the Multi-Modal
Generative Adversarial Network (mm-GAN) [10], a variant of pix2pix framework, and
trained it on intensity calibrated pre-RT and post-RT mpMRIs and the
co-registered RT dose Map (Figure 2).Results and Discussion
Co-registered
image data from an example patient in this study are shown in Figure 1: the planning
CT and associated RT dose map, and mpMRI scans acquired at three time points.
Maps of objectively-defined tissue types shown in Figure 1 provide visual
confirmation of the quality of the intensity-calibration process. Pixel
intensities on calibrated images are comparable across patients and scan dates.
The mean and SE values for pre-RT mpMRI and delta (post-RT minus pre-RT) changes
across all analyzed responding and non-responding tumors are shown in Figure 3.
The mean pre-RT ADC and T1wCE values within the GTV were higher in non-responding
tumors compared with responding tumors. Interestingly, delta T1w within the GTV
was positive in responding tumors but negative in non-responding lesions. Delta
FLAIR within the GTV was also positive in responding tumors and negative in
non-responding lesions. A similar effect was observed on delta T1wCE though the
difference was not statistically significant. Unexpectedly, there were no
significant differences on delta ADC between responding and recurrent lesions, possibly
due to the long time interval (15-158 days) between RT and post-RT MRI. Based
on these preliminary observations, the target “desirable” post-RT change within
the GTV was chosen to be +119 on delta T1w and +22 on delta FLAIR, with no
change relative to pre-RT on the other post-RT images within or outside the GTV
in a proof-of-concept demonstration of the trained mmGAN network. The predicted
and ground truth RT dose maps are compared and shown in Figure 4. The actually delivered
dose map necessarily includes dose to all tissues along the beam paths, while
this is not the case with the model predicted RT dose map.Conclusion
We
present a preliminary demonstration of a framework for training an mmGAN model
to predict SRS dose maps for voxelwise optimization of local control of BCMB
tumors. Further experiments on additional data and cohorts are required to
understand and validate this model.Acknowledgements
No acknowledgement found.References
[1] Lee H et al., "Fluence-map
generation for prostate intensity-modulated radiotherapy planning using a
deep-neural-network," Scientific Reports (2019) 9:15671.
[2] Nguyen D et al., "A feasibility
study for predicting optimal radiation therapy dose distributions of prostate
cancer patients from patient anatomy using deep learning," Scientific Reports
(2019) 9:1076.
[3] Meerbothe T, "A physics guided
neural network approach for dose prediction in automated radiation therapy
treatment planning," Master of Science thesis at the Delft University of
Technology, 2021.
[4] Ahn SH et al., "Deep learning
method for prediction of patient-specific dose distribution in breast cancer,”
Radiation Oncology (2021) 16:154.
[5] Ma M et al., "Dose distribution
prediction in isodose feature‐preserving voxelization domain using deep
convolutional neural network,” Medical Physics 46(7):2978-2987, 2019.
[6] Murakami Y et al., "Fully automated
dose prediction using generative adversarial networks in prostate cancer
patients," PLoS ONE (2020) 15(5):e0232697.
[7] Galbán CJ et al., “Diffusion MRI
in early cancer therapeutic response assessment,” NMR Biomed (2017) 30:e3458.
[8] Stringfield O et al., "Multiparameter
MRI predictors of long-term survival in glioblastoma multiforme," Tomography
(2019) 5(1):135-144.
[9] Hawkins SH et al., “MRI Predictors of
Response to Anti-PD1 Immune Checkpoint Inhibition, Bevacizumab and
Hypofractionated Stereotactic Irradiation in Patients with Recurrent High Grade
Gliomas,” Proc ISMRM, 1238, 2018.
[10] Sharma A and Hamarneh G, "Missing MRI
pulse sequence synthesis using multi-modal generative adversarial
network," IEEE Trans Med Imaging (2019) 39(4):1170-1183.