Dose-Incorporated Deep Ensemble Learning for Improving Brain Metastasis Stereotactic Radiosurgery Outcome Prediction

Abstract

Purpose/Objective(s)To develop a novel deep ensemble learning model for accurate prediction of brain metastasis(BM) local control outcomes following stereotactic radiosurgery(SRS). Materials/MethodsA total of 114 BMs from 82 patients were evaluated, including 26 BMs that developed biopsy-confirmed local failure post-SRS. The SRS spatial dose distribution(Dmap) of each BM was registered to the planning contrast-enhanced T1(T1-CE) MR. Axial slices of the Dmap, T1-CE, and PTV segmentation(PTVseg) intersecting the BM center were extracted within a fixed field-of-view determined by the V60% in Dmap. A spherical projection was implemented to transform planar image content onto a spherical surface using multiple projection centers, and the resultant T1-CE/Dmap/PTVseg projections were stacked as a 3-channel variable. Four VGG-19 deep encoders were utilized in an ensemble design, with each sub-model using a different spherical projection formula as input for BM outcome prediction. In each sub-model, clinical features after positional encoding were fused with VGG-19 deep features to generate logit results. The ensemble's outcome was synthesized from the four sub-model results via logistic regression. A total of 10 model versions with random validation sample assignments were trained to study model robustness. Performance was compared to 1) a single VGG-19 encoder; 2) an ensemble with T1-CE MRI as the sole image input after projections; and 3) an ensemble with the same image input design without clinical feature inclusion. ResultsThe ensemble model achieved an excellent AUCROC=0.89±0.02 with high sensitivity(0.82±0.05), specificity(0.84±0.11), and accuracy(0.84±0.08) results. This outperformed the MRI-only VGG-19 encoder (sensitivity:0.35±0.01, AUC:0.64±0.08), the MRI-only deep ensemble (sensitivity:0.60±0.09, AUC:0.68±0.06), and the 3-channel ensemble without clinical feature fusion (sensitivity:0.78±0.08, AUC:0.84±0.03). ConclusionFacilitated by the spherical image projection method, a deep ensemble model incorporating Dmap and clinical variables demonstrated an excellent performance in predicting BM post-SRS local failure. Our novel approach could improve other radiotherapy outcome models and warrants further evaluation.