Non-invasive CT-based Deep Learning for Human Papillomavirus Status Prediction in Oropharyngeal Cancer

Area of Science:

• Radiogenomics
• Artificial Intelligence in Oncology
• Medical Imaging Analysis

Background:

• Human papillomavirus (HPV) status is crucial for oropharyngeal cancer (OPC) treatment planning.
• Current methods for HPV classification require additional examinations.
• There is a need for non-invasive, rapid alternatives for HPV status determination.

Purpose of the Study:

• To develop a non-invasive and rapid AI-based method for HPV classification in OPC.
• To achieve performance comparable to established clinical standards.
• To integrate HPV classification into routine diagnostic and treatment workflows.

Main Methods:

• Utilized a siamese neural network with 3D DenseNet backbones.
• Incorporated tumor mask volumes and frequency-domain CT image representations.
• Trained and validated the model on 1612 cases across three independent datasets.

Main Results:

• Achieved an average AUC of 0.875 and AUPRC of 0.922.
• Demonstrated superior performance compared to reference methods across four key metrics.
• External validation showed robust generalizability with AUCs ranging from 0.726 to 0.839.

Conclusions:

• The proposed CT-based radiogenomics method achieves state-of-the-art performance for HPV prediction in OPC.
• The approach demonstrates performance comparable to existing clinical methods.
• The AI model shows "good-enough" performance for supporting patient decision-making in clinical trials, aligning with FDA guidelines.