Leveraging Artificial Intelligence and Radiomics for Improved Nasopharyngeal Carcinoma Prognostication

Abstract

Introduction

Nasopharyngeal carcinoma (NPC) typically presents as advanced disease due to the lack of significant symptoms in the early stages. Accurate prognostication is therefore challenging as current methods based on anatomical staging often lack the granularity to differentiate between patients with differing prognoses. This study investigates the potential of radiomics to improve the prediction of locoregional recurrence (LRR) and overall survival in patients with NPC.

Methods

Radiomic features were extracted from radiotherapy planning CT scans for 294 NPC patients divided into training (n = 147) and validation (n = 147) sets. A feature selection step utilising feature clustering and mutual information classifier to select six key radiomic features was employed to reduce redundancy and improve interpretability. Models were trained using clinical data, radiomic features, and these in combination to predict 2-year LRR, with performance assessed on the left-out independent validation set.

Results

Combining radiomic features with clinical data resulted in the best performance for predicting 2-year LRR (Area Under the Curve, AUC 0.76) compared to prediction using clinical or radiomic features alone (mean AUC 0.56 and 0.57, respectively). Risk stratification based on the combined model was significant for LRR-free survival and overall survival (p < 0.01). Key radiomic features included tumour size, intensity distribution, overall textural patterns, and distribution of fine and coarse textured regions.

Discussion

Radiomics holds promise for improving NPC risk stratification, potentially allowing for personalised treatment strategies. The most important radiomics feature, maximum 2D diameter, suggests a need to reconsider tumour size as a prognostic criterion despite its current exclusion from TNM staging. Larger prospective studies are needed to validate these findings.