Prediction of radiologic outcome-optimized dose plans and post-treatment magnetic resonance images: A proof-of-concept study in breast cancer brain metastases treated with stereotactic radiosurgery

Shraddha Pandey; Tugce Kutuk; Mahmoud A Abdalah; Olya Stringfield; Harshan Ravi; Matthew N Mills; Jasmine A Graham; Kujtim Latifi; Wilfrido A Moreno; Kamran A Ahmed; Natarajan Raghunand

doi:10.1016/j.phro.2024.100602

Prediction of radiologic outcome-optimized dose plans and post-treatment magnetic resonance images: A proof-of-concept study in breast cancer brain metastases treated with stereotactic radiosurgery

Phys Imaging Radiat Oncol. 2024 Jun 24:31:100602. doi: 10.1016/j.phro.2024.100602. eCollection 2024 Jul.

Authors

Shraddha Pandey^{1

2}, Tugce Kutuk³, Mahmoud A Abdalah⁴, Olya Stringfield⁴, Harshan Ravi¹, Matthew N Mills³, Jasmine A Graham^{3

5}, Kujtim Latifi^{3

5}, Wilfrido A Moreno², Kamran A Ahmed^{3

5}, Natarajan Raghunand^{1

5}

Affiliations

¹ Department of Cancer Physiology, Moffitt Cancer Center, Tampa, FL 33612, USA.
² Department of Electrical Engineering, University of South Florida, Tampa, FL 33612, USA.
³ Department of Radiation Oncology, H. Lee Moffitt Cancer Center and Research Institute, Tampa, FL 33612, USA.
⁴ Quantitative Imaging Shared Service, H. Lee Moffitt Cancer Center and Research Institute, Tampa, FL 33612, USA.
⁵ Department of Oncologic Sciences, University of South Florida, Tampa, FL 33612, USA.

Abstract

Background and purpose: Information in multiparametric Magnetic Resonance (mpMR) images is relatable to voxel-level tumor response to Radiation Treatment (RT). We have investigated a deep learning framework to predict (i) post-treatment mpMR images from pre-treatment mpMR images and the dose map ("forward models"), and, (ii) the RT dose map that will produce prescribed changes within the Gross Tumor Volume (GTV) on post-treatment mpMR images ("inverse model"), in Breast Cancer Metastases to the Brain (BCMB) treated with Stereotactic Radiosurgery (SRS).

Materials and methods: Local outcomes, planning computed tomography (CT) images, dose maps, and pre-treatment and post-treatment Apparent Diffusion Coefficient of water (ADC) maps, T1-weighted unenhanced (T1w) and contrast-enhanced (T1wCE), T2-weighted (T2w) and Fluid-Attenuated Inversion Recovery (FLAIR) mpMR images were curated from 39 BCMB patients. mpMR images were co-registered to the planning CT and intensity-calibrated. A 2D pix2pix architecture was used to train 5 forward models (ADC, T2w, FLAIR, T1w, T1wCE) and 1 inverse model on 1940 slices from 18 BCMB patients, and tested on 437 slices from another 9 BCMB patients.

Results: Root Mean Square Percent Error (RMSPE) within the GTV between predicted and ground-truth post-RT images for the 5 forward models, in 136 test slices containing GTV, were (mean ± SD) 0.12 ± 0.044 (ADC), 0.14 ± 0.066 (T2w), 0.08 ± 0.038 (T1w), 0.13 ± 0.058 (T1wCE), and 0.09 ± 0.056 (FLAIR). RMSPE within the GTV on the same 136 test slices, between the predicted and ground-truth dose maps, was 0.37 ± 0.20 for the inverse model.

Conclusions: A deep learning-based approach for radiologic outcome-optimized dose planning in SRS of BCMB has been demonstrated.

Keywords: Brain metastases; Deep learning; Dose map prediction; Dose painting; Image prediction; Stereotactic radiosurgery.