Author: Stephen R. Bowen, Shijun Chen, Chunyan Duan, Daniel S. Hippe, Qiantuo Liu, Qianqian Tong, Jiajie Wang, Shouyi Wang, Faisal Yaseen 👨🔬
Affiliation: The University of Texas at Austin, Tongji University, University of Washington, Department of Radiation Oncology, Fred Hutchinson Cancer Center, University of Washington, Fred Hutchinson Cancer Center, University of Texas at Arlington 🌍
Purpose: Tumor subregion clustering and prediction of region-specific response can augment assessments and adaptive treatment decisions. A modeling framework was constructed to predict chemoradiation response in tumor subregions on FDG PET.
Methods: 23 locally advanced non-small cell lung cancer patients enrolled on the FLARE-RT trial (NCT02773238) underwent FDG-PET/CT imaging prior to (PrePET) and during week 3 (MidPET) of chemoradiotherapy. Tumors were clustered into high-risk and low-risk subregions in two stages, using voxel position and standardized uptake value (PrePET, MidPET, or both). Each feature was weighted in the clustering, and the choice of weights was optimized in each stage to maximize clustering metrics. 3 conventional features and 41 radiomics features were extracted from high-risk tumor subregions, low-risk tumor subregions and the whole tumor to perform regression prediction(∆SUVmean=[SUVmean_mid–SUVmean_pre]/SUVmean_pre) and classification prediction(Response class=[≥20% decrease in ∆SUVmean]) of mid-chemoradiation response based on auto machine learning. Silhouette coefficient (SC), Calinski-Harabasz index (CH), dice similarity (DS), and Euclidean distance (ED) metrics were used to evaluate the clustering performance. The area under ROC curve (AUC) was used to evaluate classification prediction performance and root mean square error(RMSE) was used to evaluate regression prediction performance, following leave-one-patient-out cross-validation.
Results: Two-stage clustering achieved similar or improved SC and CH, as well as higher DS (DS_high=0.91 and DS_low=0.93) and lowest ED (ED_high=0.54 and ED_low=0.42) compared with other clustering algorithms. For auto machine learning, AUC (3all=0.74, 3high=0.63 and 3low=0.67; 41all=0.60, 41high=0.67 and 41low=0.60) were achieved for classification prediction. RMSE (3all=0.19, 3high=0.17 and 3low=0.19; 41all=0.17, 41high=0.19 and 41low=0.22) were achieved for regression prediction. Performance was consistently numerically better than performance achieved by other common machine learning algorithms (AUC:0.33-0.59, RMSE:0.17-0.29).
Conclusion: Two-stage clustering and machine learning can predict chemoradiation response in tumor subregions with good performance for both continuous and binary response measures. This framework can support treatment assessment and adaptive therapies.