Author: Evan Calabrese, Hangjie Ji, Kyle J. Lafata, Casey Y. Lee, Eugene Vaios, Chunhao Wang, Lana Wang, Zhenyu Yang, Jingtong Zhao 👨🔬
Affiliation: Duke University, Department of Radiation Oncology, Duke University, Duke Kunshan University, North Carolina State University 🌍
Purpose: To develop a biologically guided deep learning (DL) model for predicting brain metastasis(BM) local control outcomes following stereotactic radiosurgery (SRS). By integrating pre-SRS MR images and spatial dose distributions into a neural partial differential equation (NPDE) framework, the model characterizes tumor responses to radiation across temporal and spatial dimensions.
Methods: By hypothesizing that changes in MR in response to radiation therapy can be described using a biology-based mathematical model, we implemented a NPDE framework governed by a reaction-diffusion mechanism. Patient-specific spatial dose distribution was incorporated to reflect treatment effects. A U-shape deep neural network (DNN) was designed and trained to learn parameters of the biological model. This enabled visualization of tumor biological dynamics, including diffusion, proliferation, and treatment response, by solving the PDE. The model generated time-series predictions of post-SRS MR imaging states from pre-SRS inputs, and then ensembled intermediate biological states to predict BM outcomes. The dataset included 128 BMs from 96 patients with paired 3D pre- and post-SRS MR scans, including 32 BMs that developed biopsy-confirmed post-SRS local failure. The proposed two-branch DNN simultaneously learned biological model terms from paired MR images and spatial dose distributions in one branch while generating post-SRS image predictions in the other. An 8:2 train-test split was applied, and ten model versions were trained to ensure robustness. Sensitivity, specificity, accuracy, and ROCAUC were used for evaluation.
Results: The model achieved an ROCAUC of 0.82±0.05, sensitivity of 0.75±0.17, specificity of 0.81±0.09, and accuracy of 0.77±0.11. Breakdown analysis of biological dynamics reveals spatial retraction of abnormal metabolism and effective suppression of malignancy over time. The predicted post-SRS MR images aligned well with clinical expectations.
Conclusion: The proposed biologically guided DL method effectively predicts BM post-SRS outcomes. By incorporating biological components, the predicted imaging outcomes can support adaptive radiotherapy strategies, enabling optimized, patient-specific BM management.