Radiomics Model Using CT to Predict Progression in Locally Advanced Head and Neck Cancer
Overview
A combined radiomics and clinicobiological model was developed to predict progression-free survival (PFS) in locally advanced head and neck squamous cell carcinoma (LAHNSCC). This CT-based model, enhanced by SHAP explainability, aims to outperform traditional TNM8 and HPV status stratification methods.
Background
Head and neck squamous cell carcinoma (HNSCC) is the seventh most common cancer globally, with over 60% presenting as locally advanced disease. Standard treatment involves high-dose cisplatin with radiotherapy or cetuximab in unfit patients, yet recurrence remains high. Current prognostic tools like TNM8 staging and HPV status are insufficient for precise risk stratification. Radiomics, which extracts quantitative imaging features from CT scans, offers potential to improve prognostic accuracy but has faced challenges in validation and clinical implementation.
Data Highlights
The study included LAHNSCC patients treated from 2016 to 2023 at a single institution, with CT scans acquired within 28 days before treatment. Radiomics features were extracted from 3D tumor volumes segmented on contrast-enhanced CT images. Clinical variables collected included age, sex, tumor location, TNM8 stage, performance status, tobacco and alcohol use, and HPV status. Progression-free survival (PFS) was the primary outcome, defined from diagnosis to progression or last follow-up.
Key Findings
A combined model integrating CT radiomics features with clinical and biological data predicted PFS more accurately than TNM8 staging and HPV status alone.
Use of SHAP values provided interpretable insights into feature contributions, enhancing clinical relevance and decision-making transparency.
Contrast-enhanced CT imaging and manual 3D tumor segmentation enabled extraction of robust radiomics features compliant with IBSI standards.
The model was developed and validated on a well-defined, single-institution cohort with standardized imaging protocols, reducing variability.
Radiomics features captured tumor heterogeneity beyond visual assessment, potentially identifying subvisual markers of progression risk.
Clinical Implications
Incorporating radiomics with clinical and biological variables can improve risk stratification for LAHNSCC patients, potentially guiding personalized treatment decisions. The explainability provided by SHAP values may facilitate clinician acceptance and integration of AI-driven models into routine practice. Standardized imaging acquisition and rigorous segmentation protocols are critical for reproducible radiomics analysis.
Conclusion
This study demonstrates that a CT-based radiomics model combined with clinicobiological data can enhance prediction of disease progression in LAHNSCC, offering a promising tool to complement existing staging systems. Further validation and implementation efforts are warranted to translate these findings into clinical workflows.
References
Global Cancer Statistics 2020 -- Head and Neck Cancer Prevalence
