Comparation of CT/PET/MRI in deep learning models for outcome prediction in oropharyngeal cancer patients

Background Due to the superior soft tissue contrast, MRI may provide more prognostic information than CT/PET for outcome prediction in oropharyngeal cancer (OPC). However, the utilization of MRI-based deep learning (DL) models to predict outcome of OPC patients is rarely explored. Purpose This study aims to compare the prognostic value of MRI with CT and PET in DL models for local control (LC), regional control (RC) and overall survival (OS) in OPC patients in relation to a clinical benchmark model. Methods A dataset comprising 266 OPC patients who received (chemo)radiotherapy was assembled. Each patient’s data includes a pretreatment axial T1-weighted scan (T1), a coronal T2-weighted MRI scan (T2), CT and PET scans, contoured Gross Tumor Volume (GTV) of the primary tumor, clinical parameters and follow-up information regarding LC, RC and OS. Patients were divided into a training set (n = 186) and a test set (n = 80). Various 2D and 3D convolutional neural networks (CNNs) were trained using CT, PET, T1 or T2 images of the contoured GTV volume with and without a margin for outcome prediction. Results The 2D models using T2 images within a bounding box region determined by the GTV volume with a 5-mm margin achieved a notably high concordance index (C-index) of 0.88 and 0.75 for LC and OS prediction, respectively. Additionally, MRI-based models achieved much higher C-indexes than CT- or PET-based models for LC prediction. In comparison to a clinical benchmark model (HNC-PREDICTOR), the T2-based 2D model showed an improved LC prediction (C-index: 0.88 vs. 0.80) and combining the T2 prediction and clinical model resulted in an improved OS prediction (C-index 0.81 vs. 0.78). Furthermore, the clinical + MRI models showed enhanced performance compared to a clinical routine OS risk stratification system and good calibration. Conclusions MRI-based DL models can improve LC and OS prediction of OPC compared to CT, PET-based and clinical benchmark models.