Early prediction of radiotherapy outcomes in pharyngeal cancer using deep learning on baseline [18F]Fluorodeoxyglucose positron emission Tomography/Computed tomography

Abstract

Objectives

This study aimed to develop an integrated segmentation-free deep learning (DL) framework to predict multiple aspects of radiotherapy outcome in pharyngeal cancer patients by analyzing pretreatment 18F-fluorodeoxyglucose (18F-FDG) positron emission tomography-computed tomography (PET/CT).

Methods

We utilized baseline 18F-FDG-PET/CT scans from patients newly diagnosed with oropharyngeal or hypopharyngeal cancer. The study cohort comprised 162 patients for training and 32 for validation, all of whom completed definitive chemoradiotherapy or radiotherapy for organ-preservation. Following image augmentation, fused PET and CT images were used to train three distinct DL models. An ensemble voting classifier was then employed to predict local recurrence (LR), neck lymph node relapse (NR), and distant metastases (DM). Model performance was evaluated using receiver operating characteristic curve analysis.

Results

With a median follow-up of 36 months, the training cohort experienced, LR in 45 (27.8 %), NR in 32 (19.8 %), and DM in 21 (13.0 %) patients. By optimizing single models and finalizing with an ensemble voting classifier, the area under the curve for the occurrence of LR, NR, and DM was 0.850, 0.878, and 0.893, whereas the accuracy for the three endpoints were 87.5 %, 68.8 %, and 78.1 %, respectively.

Conclusions

By utilizing baseline 18F-FDG-PET/CT, our proposed DL models can provide a supplemental prediction for various therapeutic outcome in patients with pharyngeal cancer undergoing radiotherapy-based treatment. The accuracy for NR and DM predictions requires further optimization through additional technological breakthrough or combing clinical parameters. External validation is an important future step to confirm the model’s generalizability and clinical utility.