Predicting taste alteration in head and neck cancer patients using 3D transformer-based deep learning model

Taste alteration is a common post-treatment side effect of head and neck cancer (HNC) radiotherapy. Accurate normal tissue complication probability (NTCP) models can support treatment planning aimed at reducing side effects that impair quality-of-life. Deep learning (DL) offers opportunities to incorporate full 3D dose, CT, and organofinterest (OOI) data. Using a large prospective cohort, this study explores whether CNN and transformerbased DL models can outperform a conventional logistic regression model for predicting taste alteration 6 months after radiotherapy.

Data from 972 patients was used, with the endpoint being moderatetosevere patientreported taste alteration at six months (EORTC QLQH&N35). DL models were trained using 3D dose distributions, planning CT, OOI segmentations and clinical features. A conventional NTCP model using oral cavity mean dose, parotid mean dose, and age served as reference. Attention maps were evaluated for the transformer-based DL model to quantify the model attention to specific regions of the head and neck area. In addition, two cropping approaches for the 3D data were compared.

The CNN and transformer DL models performed comparable to the reference NTCP model in the test set (Reference: AUC = 0.71; ResNet: AUC = 0.68; TransRP: AUC = 0.71), though calibration was best for the reference model. Shifting the crop improved DL performance and transformer attention focused on supraglottic and submandibular regions.

This study showed that the prediction of taste alteration can be achieved using DL models, particularly transformer-based models which, in contrast to CNNs, can account for global features of comprehensive 3D information of dose distribution, CT-images and organs-of-interest segmentations.