Radiation therapy response prediction for head and neck cancer using multimodal imaging and multiview dynamic graph...

Area of Science:

• Medical Imaging
• Radiomics
• Cancer Biomarkers

Background:

• External beam radiation therapy is a standard treatment for head and neck (H&N) cancers.
• Radiomic features from medical images show potential as biomarkers for tumor heterogeneity and treatment response prediction.
• Current methods often use single imaging modalities or simple concatenation of features, limiting comprehensive analysis.

Purpose of the Study:

• To evaluate multiview feature selection (MVFS) for identifying key radiomic features from quantitative ultrasound spectroscopic (QUS) parametric maps, computed tomography (CT), and magnetic resonance imaging (MRI).
• To train predictive models using selected features to forecast radiation therapy outcomes in H&N cancer patients.

Main Methods:

• Extracted 70, 70, and 350 radiomic features from CT, MRI, and QUS parametric maps, respectively.
• Developed an Adaptive Graph Autoencoder Multi-View Feature Selection (AGAMVFS) technique utilizing dynamic graph learning and autoencoders for feature discrimination and selection.
• Employed leave-one-patient-out cross-validation and trained support vector machine (SVM) and k-nearest neighbor (KNN) classifiers to predict treatment response.

Main Results:

• The AGAMVFS method combined with an SVM classifier achieved 85% accuracy, 76% sensitivity, 91% specificity, and 83% balanced accuracy in predicting H&N cancer treatment response.
• The proposed method demonstrated superior performance compared to single-modality approaches and existing feature selection techniques.
• Top selected features included six QUS, three MRI, and one CT radiomic feature, highlighting the value of integrating multiple imaging data.

Conclusions:

• The developed predictive model effectively predicts treatment response for H&N cancer patients.
• MVFS enhances feature interpretability and maintains inter-correlations among features from diverse imaging modalities, offering a more robust analytical framework.