Deep Learning Based Prognostic Prediction in Oropharyngeal Cancer Patients Using Multiparametric MRI Inputs

Abstract

While prognostic outcomes for oropharyngeal cancer (OPC) patients have improved in recent years, patients still face a non-negligible risk of disease recurrence or death. Accurately predicting post-therapy prognosis would be highly valuable for risk stratification and treatment guidance for OPC patients. Recent studies using PET/CT data have demonstrated the effectiveness of large-scale, end-to-end image-based deep learning (DL) models for predicting progression-free survival (PFS) in OPC patients. Multiparametric MRI (mpMRI), which combines anatomical and functional MRI sequences, has the potential to offer similar results, and has the added advantage of high-frequency longitudinal imaging capabilities, such as through MR-Linac devices. Therefore, this study aimed to develop a DL model using mpMRI data to predict PFS, and to evaluate the impact of anatomical and functional input channels on model performance. From a large-scale head and neck cancer database at MD Anderson Cancer Center, treatment-naïve OPC patients with available pre-radiotherapy mpMRI imaging were selected for this study. mpMRI images used for this study included T2-weighted images (T2) and apparent diffusion coefficient (ADC) maps. PFS event status was defined as having either a local, regional, or distant failure, and/or death; data were right censored if an event had not occurred. Images were resampled to the T2 resolution, normalized to a [-1,1] scale, and cropped to the field of view of the ADC image for use in DL models. A DL convolutional neural network model based on the DenseNet121 architecture from the Medical Open Network for AI (MONAI) Python package using a negative log-likelihood loss function was implemented. The model used mpMRI images as input channels and 20 output channels representing the different time intervals of the predicted PFS conditional probabilities of surviving that time interval; final PFS in days was obtained by summing the cumulative probability of surviving each interval times the interval duration. A 5-fold cross validation approach was used for model training and evaluation. Separate models using only T2, only ADC, and T2 + ADC channel inputs were compared. Model performance was measured using the C-index. Out of 1154 patients, 404 met inclusion criteria. The overall PFS event rate was 16%. Median C-index values from the 5-fold cross validation were 0.62, 0.67, and 0.69 for the ADC, T2, and T2+ADC models, respectively. Using large-scale datasets and open-source DL implementations, we find that OPC PFS prediction models using mpMRI data yield modest but comparable performance to existing models (i.e., state-of-the-art reference performance using PET/CT). Moreover, combining mpMRI channels may increase the performance of models for OPC prognostic prediction. Future work will involve integration of additional timepoints, additional mpMRI images, clinical variables, and saliency maps.