Feature level quantitative ultrasound and CT information fusion to predict the outcome of head & neck cancer radiotherapy treatment: Enhanced principal component analysis

Abstract Background Radiation therapy is a common treatment for head and neck (H&N) cancers. Radiomic features, which are determined from biomedical imaging, can be effective biomarkers used to assess tumor heterogeneity and have been used to predict response to treatment. However, most studies employ only a single biomedical imaging modality to determine radiomic features. Purpose The objective of this study was to evaluate the effectiveness of radiomic feature fusion, combining quantitative ultrasound spectroscopy (QUS) and computed tomography (CT) imaging modalities, in predicting the outcomes of radiation therapy for H&N cancer prior to start. Method An enhanced version of principal component analysis (EPCA) was proposed to fuse 70 radiomic features from CT and 476 radiomic features from QUS in order to predict the response to radiation therapy in patients with H&N cancers (partial response vs. complete response). EPCA is a PCA method with Hessian matrix regularization andl 2 , 1 − 2 ${{l}_{2,1 - 2}}$ ‐regularization, and was proposed here for information fusion at a feature level. Leave‐one‐patient‐out methodology with bootstrap was applied to conduct train‐test analysis and fused features were used to train two (support vector machine (SVM) and k‐nearest neighbor (KNN)) classifiers to build a predictive model in order to predict response to treatment for patients with H&N cancers. Five‐fold (5) cross validation was applied on the training set to tune the hyperparameters of SVM and KNN classifiers. Consequently, the performance of classifiers was evaluated by examining accuracy (ACC), F1‐score (F1), balanced accuracy (BACC), Sensitivity (S n ), and Specificity ( S p ) metrics. Additionally, a two‐sided t ‐test was applied to the top principal components derived from EPCA methodology in order to assess the statistical significance of the selected components. The proposed method developed here was compared with minimum redundancy maximum relevance (mRMR) feature selection, conventional PCA, kernel PCA, autoencoder, and canonical correlation analysis (CCA). Additionally, we compared proposed EPCA with robust PCA andl 2 , 1 ${{l}_{2,1}}$ ‐norm constrained graph Laplacian PCA. Results Seventy‐one ( n  = 71) (66 male (93%) and female (7chmch%)) H&N cancer patients were recruited with bulky metastatic neck lymph node (LN) involvement. Patients had a mean age of 59 ± 10 and 25 (35.2%) were complete responders and 46 (64.8%) were partial‐responders. In terms of predicting responses, the EPCA‐SVM classifier had better performance than EPCA‐KNN, and achieved 79± $\ \pm $ 2% sensitivity, 84± $\ \pm $ 2% specificity, 82  ± $ \pm $ 2% accuracy, 81± $\ \pm $  2% balanced accuracy, and 82  ± 2 $ \pm \ 2$ % area under curve (AUC). Results demonstrated the effectiveness of the proposed method with superiority over mRMR feature selection, conventional PCA, kernel PCA, autoencoder, and CCA methods. Using an ablation study, EPCA was compared with robust PCA andl 2 , 1 ${{l}_{2,1}}$ ‐norm constrained graph Laplacian PCA. Results supported the superiority of EPCA over rPCA andl 2 , 1 ${{l}_{2,1}}$ ‐norm constrained graph Laplacian PCA. Three principal components were statistically significant. Additionally, we compared the proposed method with the use of QUS and CT as individual imaging modalities. The results demonstrated the effectiveness of feature‐level fusion in enhancing prediction accuracy. Conclusion The results demonstrated that the proposed predictive model is able to predict a binary H&N cancer treatment outcome, feature level fusion of CT and QUS radiomics has superiority over single imaging modality and EPCA is an effective approach to fuse the features.