Imaging-based Prognostic Artificial Intelligence Model for Oropharyngeal Carcinoma after Radiation Therapy
Recommended Citation
Zhu S, Gilbert MV, Liu P, Siddiqui F. Imaging-based Prognostic Artificial Intelligence Model for Oropharyngeal Carcinoma after Radiation Therapy. Int J Radiat Oncol Biol Phys 2024; 118(5):e64.
Document Type
Conference Proceeding
Publication Date
4-1-2024
Publication Title
Int J Radiat Oncol Biol Phys
Abstract
Purpose/Objective(s): Predicting treatment prognosis for oropharyngeal carcinoma (OPC) remains a significant challenge. This study hypothesizes that pre-treatment CT imaging holds important prognostic information. We developed an artificial intelligence (AI) model to predict locoregional recurrence (LRR) post-radiation therapy (RT) for OPC. Materials/Methods: In an IRB-approved study, we collected imaging and outcome data from 1095 patients newly diagnosed with oropharyngeal carcinoma. They were treated with RT with curative intent, with or without chemotherapy, and none underwent surgery. Data originated from multiple institutions: 124 from ours and 971 from The Cancer Imaging Archive, contributed by four institutions in the US, Canada, and Europe. We excluded patients with a follow-up duration of less than 2 years unless they experienced LRR within this period. Each patient's pre-treatment CT images, along with segmentations for the gross tumor volume of the primary tumor and lymph nodes (if present), served as model input. We clipped the Hounsfield values of CT images within the range [-200, 200]. A 3D convolutional neural network, adapted from the ConvNeXt architecture, was employed as the deep learning model. The primary endpoint was the risk of 2-year LRR. The training used a weighted binary cross-entropy loss function, and we developed the model through five-fold cross-validation on 730 cases. For inference, the ensemble model provided a risk score between 0 and 1, indicating the probability of 2-year LRR on the 365 patients in the test cohort. Results: On the independent test set, the model attained a Harrel's concordance index of 0.72 for predicting 2-year LRR. Using 0.5 as a threshold to classify the test cohort into low- and high-risk groups, the log-rank test highlighted a significant difference in LRR risk between the two groups (p=3.6 × 10^-7). Conclusion: Our study demonstrates the potential of an imaging-based AI model for predicting OPC prognosis. However, further research is required to validate this model and integrate more clinical parameters as inputs.
Volume
118
Issue
5
First Page
e64
