2483 - Spatial-Dose Attention Using Explainable AI: A Deep Learning Framework to Predict Complete Obliteration in AVM Radiosurgery
Presenter(s)
P. W. Huang1,2, W. Y. Chung3, D. H. C. Pan3, C. C. Lee4, H. C. Yang4, M. C. Chen5, J. R. Jhuang6, J. T. Tsai2, and L. J. Chen2; 1Graduate Institute of Clinical Medicine, College of Medicine, Taipei Medical University, Taipei, Taiwan, 2Department of Radiation Oncology, Shuang Ho Hospital, Taipei Medical University, New Taipei City, Taiwan, 3Department of Neurosurgery, Shuang Ho Hospital, Taipei Medical University, New Taipei City, Taiwan, 4Department of Neurosurgery, Taipei Veterans General Hospital, Taipei, Taiwan, 5Department of Biomedical Imaging and Radiological Sciences, National Yang Ming Chiao Tung University, Taipei, Taiwan, 6Institute of Epidemiology and Preventive Medicine, College of Public Health, National Taiwan University, Taipei, Taiwan
Purpose/Objective(s): Predicting complete obliteration (CO) after arteriovenous malformation (AVM) radiosurgery (SRS) remains challenging. Global dose metrics often fail because dose-response effects dilute across non-critical nidus areas. We hypothesized that localized under-dosing within specific high-risk subregions drives treatment failure. We aimed to develop an interpretable deep learning framework using explainable AI (XAI) to integrate MRI and 3D dosimetry, localizing these critical targets via a spatial-dose attention mechanism.
Materials/Methods: We retrospectively analyzed 285 unruptured AVM patients (207 CO, 78 non-CO) treated with SRS across two centers. T2-weighted MRI contours were isotropically expanded 1.5 cm to capture parenchymal and hemodynamic context. We built a two-stage deep learning pipeline. Stage 1 trained MRI-only CNN models for CO classification using focal loss, applying Grad-CAM to extract high-importance spatial features. Stage 2 introduced the dose-attention mechanism: Grad-CAM heatmaps were thresholded (>0.75) to define critical subregions. After strict spatial co-registration of MRI, 3D dose, and Grad-CAM maps, voxel-wise doses within these boundaries were extracted. Rather than naive channel concatenation, we utilized a dual-channel input (MRI+dose) with learned attention for end-to-end spatial dose weighting. All images underwent percentile-based intensity normalization. Performance was evaluated via 5-fold cross-validation (65/15/20 splits) following intensity normalization.
Results: The baseline MRI-only model yielded an F1 score of 0.87 (accuracy 0.78, sensitivity 0.80, specificity 0.97). Grad-CAM localized critical features to the nidal arterial end, corroborating the hemodynamic hypothesis. Integrating radiation parameters via simple concatenation (MRI+Dose) hyper-sensitized the model (sensitivity 0.99) but severely compromised specificity (0.54), dropping accuracy to 0.72 and F1 score to 0.84. Conversely, our dose-attention model, which quantifies voxel-level dose strictly within XAI-defined critical subregions, achieved the highest F1 score (0.89) and accuracy (0.83). It maintained high sensitivity (0.96) while successfully rescuing specificity (0.75), demonstrating discriminative superiority over naive concatenation.
Conclusion: Global dose metrics dilute the true dose-response required for AVM obliteration. Our XAI-driven spatial-dose attention framework successfully mapped critical functional targets—approximating the hypothesized arterial feeding zones—proving that spatially weighted dosimetry statistically outperforms both MRI-alone and naively concatenated models. Clinically, this interpretable pipeline delivers an actionable decision-support tool to pinpoint under-dosed high-risk nidus compartments, enabling clinicians to prospectively optimize SRS treatment plans and maximize CO rates.