Multi-Modal Ensemble Deep Learning in Head and Neck Cancer HPV Sub-Typing

Area of Science:

• Radiology and Oncology
• Artificial Intelligence in Medicine
• Medical Imaging Analysis

Background:

• Oropharyngeal Squamous Cell Carcinoma (OPSCC) exhibits significant heterogeneity, with human papillomavirus (HPV) infection being a key risk factor.
• Accurate HPV status determination in OPSCC is crucial for guiding treatment strategies and predicting patient outcomes.
• Current diagnostic methods may require invasive procedures, highlighting the need for non-invasive alternatives.

Purpose of the Study:

• To develop and evaluate a deep learning-based method for automated HPV status detection in OPSCC.
• To assess the accuracy of a multi-modal feature fusion architecture using Computed Tomography (CT) and Positron Emission Tomography (PET) images.
• To compare the performance of the proposed model against existing methods for HPV status classification.

Main Methods:

• A 3D Convolutional Neural Network (CNN)-based multi-modal feature fusion architecture was designed for HPV status prediction.
• The model integrated features related to intensity, contrast, shape, texture heterogeneity (from CT), and metabolic assessment (from PET) of tumor regions.
• An ensemble of CNN networks with soft voting was employed to merge multi-modal features for final classification.

Main Results:

• The multi-modal ensemble model demonstrated superior performance compared to single-modality PET/CT approaches.
• The proposed method achieved an Area Under the Curve (AUC) of 0.76 and an F1 score of 0.746 on combined TCGA and MAASTRO datasets.
• On the MAASTRO dataset, the model attained an AUC score of 0.74 for primary tumor volumes of interest (VOIs).

Conclusions:

• Deep learning-based analysis of CT and PET images can accurately predict HPV status in OPSCC.
• The multi-modal feature fusion approach offers a promising non-invasive tool for OPSCC diagnosis.
• Further validation in larger cohorts is recommended to enhance diagnostic accuracy and support pre-biopsy assessments.