Do Spatial-Radiomics Improve Prediction of Locoregional Recurrence Following Radiotherapy for HNSCC?

Purpose/Objective(s) Definitive chemoradiation for HNSCC has improved significantly with modern planning techniques and supportive care. However the locoregional recurrence risk (LR) persists and can range up to 30% depending on patient-specific factors. Prior research has used machine learning on radiomic features (or quantitative, sub-visual cues) from diagnostic imaging to predict LR with some success. In this study, we attempt to improve on these models by introducing 2 key innovations. First we identify "supervoxels," or sub-regions near but outside of the gross tumor volume (GTV), from which to extract radiomic features. Second we create a radiomic graph model where the supervoxels are prioritized based on their similarity to the GTV. We hypothesize that this spatial, graph-based approach can better identify regions suspicious for microscopic tumor involvement, which in turn would better predict clinical outcomes. Materials/Methods We identified the RADCURE (D1) and Head-Neck-Radiomics-HN1 (D2) datasets containing 2,611 patient CTs and RT structures from The Cancer Imaging Archive. D1 was divided into training, validation, and testing splits whereas D2 was used only for testing. For each patient CT, we identified 100 supervoxels within 10 voxels of the GTV. Texture-based radiomic features were then extracted from each supervoxel, and the top 20 with feature expression most similar (via Euclidean distance) to the GTV used to create a graph model. The nodes of this graph represent feature expression within the supervoxels and the edges the correlation to the GTV. Clinical features including age, sex, ECOG, chemotherapy use, tumor stage, size, and HPV status were used as a model parameter and baseline comparison. A graph attention neural network (GAT) was trained using these graphs for the LR prediction task. Comparisons were made with a traditional radiomic model and clinical features alone. Following prediction, model attention weights were extracted and used to identify which CT supervoxels were most informative to the model. Results Graph radiomics with clinical features resulted in AUCs of 0.834 and 0.806 for D1 and D2, respectively. Traditional radiomics with clinical features resulted in AUCs of 0.819 and 0.784 compared to clinical features alone achieving AUCs of 0.808 and 0.784. Qualitative examination of attention heatmaps revealed that our spatial radiomic model attention was heavily concentrated along cervical lymph node chains. Conclusion Spatial radiomics utilizing supervoxels from peritumoral areas were able to predict LR for HNSCC in large, multi-institutional datasets, outperforming other previously studied methods. It is notable that our model's performance did indeed improve on an already robust baseline for an independent test dataset, suggesting there is additional utility in our graph-based approach. Our attention maps further suggest that disease-relevant regions outside of the GTV can be identified in an unsupervised manner. Definitive chemoradiation for HNSCC has improved significantly with modern planning techniques and supportive care. However the locoregional recurrence risk (LR) persists and can range up to 30% depending on patient-specific factors. Prior research has used machine learning on radiomic features (or quantitative, sub-visual cues) from diagnostic imaging to predict LR with some success. In this study, we attempt to improve on these models by introducing 2 key innovations. First we identify "supervoxels," or sub-regions near but outside of the gross tumor volume (GTV), from which to extract radiomic features. Second we create a radiomic graph model where the supervoxels are prioritized based on their similarity to the GTV. We hypothesize that this spatial, graph-based approach can better identify regions suspicious for microscopic tumor involvement, which in turn would better predict clinical outcomes. We identified the RADCURE (D1) and Head-Neck-Radiomics-HN1 (D2) datasets containing 2,611 patient CTs and RT structures from The Cancer Imaging Archive. D1 was divided into training, validation, and testing splits whereas D2 was used only for testing. For each patient CT, we identified 100 supervoxels within 10 voxels of the GTV. Texture-based radiomic features were then extracted from each supervoxel, and the top 20 with feature expression most similar (via Euclidean distance) to the GTV used to create a graph model. The nodes of this graph represent feature expression within the supervoxels and the edges the correlation to the GTV. Clinical features including age, sex, ECOG, chemotherapy use, tumor stage, size, and HPV status were used as a model parameter and baseline comparison. A graph attention neural network (GAT) was trained using these graphs for the LR prediction task. Comparisons were made with a traditional radiomic model and clinical features alone. Following prediction, model attention weights were extracted and used to identify which CT supervoxels were most informative to the model. Graph radiomics with clinical features resulted in AUCs of 0.834 and 0.806 for D1 and D2, respectively. Traditional radiomics with clinical features resulted in AUCs of 0.819 and 0.784 compared to clinical features alone achieving AUCs of 0.808 and 0.784. Qualitative examination of attention heatmaps revealed that our spatial radiomic model attention was heavily concentrated along cervical lymph node chains. Spatial radiomics utilizing supervoxels from peritumoral areas were able to predict LR for HNSCC in large, multi-institutional datasets, outperforming other previously studied methods. It is notable that our model's performance did indeed improve on an already robust baseline for an independent test dataset, suggesting there is additional utility in our graph-based approach. Our attention maps further suggest that disease-relevant regions outside of the GTV can be identified in an unsupervised manner.