Radiation therapy response prediction for head and neck cancer using multimodal imaging and multiview dynamic graph autoencoder feature selection

Background

External beam radiation therapy is a common treatment for head and neck (H&N) cancers. Radiomic features derived from biomedical images have shown promise as effective biomarkers used to assess tumor heterogeneity and predict response to treatment. However, most studies employ only a single biomedical imaging modality to determine radiomic features or naively concatenate radiomic features from different imaging modalities.

Purpose

The objective of this study is to assess the effectiveness of multiview feature selection (MVFS) in identifying the most discriminative radiomic features determined from pretreatment quantitative ultrasound spectroscopic (QUS) parametric maps, as well as computed tomography (CT), and magnetic resonance imaging (MRI) modalities. These features were used to train predictive models to predict outcomes of radiation therapy for head and neck (H&N) cancer.

Method

70, 70, and 350 radiomics features were extracted from pre-treatment CT and MRI images, as well as seven QUS parametric maps, respectively. We proposed an MVFS technique named Adaptive Graph Autoencoder Multi-View Feature Selection (AGAMVFS), based on dynamic graph learning and autoencoder. In AGAMVFS, adaptive and collaborative graphs are learned at multiple levels to discriminate among view-specific features. An autoencoder is then applied to concatenated features to select the most discriminative ones. This approach fosters collaboration across different views and achieves a consensus projection for feature selection. Leave-one-patient-out cross-validation was applied to split the data into train and test sets and selected features were used to train two classifiers (support vector machine (SVM) and k-nearest neighbor (KNN)) to build a predictive model, tasked with predicting response to treatment for patients with H&N cancers. Fivefold cross-validation was applied on training set to tune the hyperparameters of SVM and KNN classifiers. Consequently, the performance of classifiers was evaluated using accuracy, F1-score, balanced accuracy, sensitivity, and specificity metrics. Additionally, a two-sided t-test was applied to the selected features. We compared the proposed method with a single imaging modality and state-of-the-art feature selection techniques.

Results

We recruited 63 (59 male (94%) and 4 female (6%)) H&N cancer patients with bulky metastatic neck lymph node (LN) involvement. The mean age was 58.9 ± 10.2 years. The AGAMVFS with the SVM classifier obtained the best performance and achieved 76% sensitivity, 91% specificity, 85% accuracy, and 83% balanced accuracy. Results showed the effectiveness of proposed method with superiority over other feature selection techniques. The most top-10 frequent features were six QUS radiomics, three MRI radiomics, and one CT radiomics features.

Conclusion

The results demonstrated that the proposed predictive model is able to predict H&N cancer treatment response. MVFS provided better interpretabilityfor analysing features and preserved the inter-correlation among features from different imaging modalities.