Classification of chemoradiotherapy sensitivity in hypopharyngeal squamous cell carcinoma based on deep-learning and radiomics feature fusion.

Abstract

Background: Analyzing apparent diffusion coefficient (ADC) images before chemoradiotherapy (CRT) can effectively predict the treatment response of patients with hypopharyngeal squamous cell carcinoma (HPSCC), thereby reducing the treatment risks. This study aimed to develop a predictive model by combining deep-learning features and radiomics features derived from ADC images to predict the CRT sensitivity of HPSCC patients, providing effective guidance for treatment strategy selection.

Methods: This study retrospectively analyzed the data of 120 HPSCC patients. Deep-learning features were extracted from ADC images using a vision transformer (ViT)-based deep-learning model, while radiomics features were extracted using the PyRadiomics feature extractor. Among the 1,288 extracted radiomics features, the most significant ones were selected using the Spearman's correlation coefficient, intraclass correlation coefficient (ICC), and least absolute shrinkage and selection operator (LASSO) method. These features were fused through a concatenation approach, and a classification prediction was performed using a convolutional neural network with three fully connected layers.

Results: The accuracy, sensitivity, specificity, and area under the curve (AUC) values of the feature fusion model were 0.99 vs. 0.875, 0.988 vs. 0.842, 1.000 vs. 1.000, and 1.000 vs. 0.947, for the training and validation datasets, respectively. The feature fusion model performed optimally in comparison to the other models. In the validation dataset, the accuracy of the feature fusion model improved by 16.7% and 4.2% compared to the clinical and radiomic models, respectively.

Conclusions: The model developed in this study, which integrates deep-learning features with traditional radiomics features, can accurately predict the CRT sensitivity of HPSCC patients using pre-treatment ADC images. This model provides an effective reference for selecting optimal treatment strategies for patients.