Machine Learning-Based Classification of Cervical Lymph Nodes in HNSCC: A Radiomics Approach with Feature Selection...

Area of Science:

• Radiology
• Oncology
• Medical Imaging Analysis

Background:

• Computed tomography (CT) is vital for head and neck squamous cell carcinoma (HNSCC) staging, but conventional lymph node (LN) evaluation lacks sensitivity.
• Morphological criteria in CT are subjective and may miss early metastatic disease, impacting prognosis and treatment planning.

Purpose of the Study:

• To leverage radiomics for quantitative feature extraction from CT images to improve LN classification accuracy in HNSCC.
• To develop a noninvasive method for LN assessment, reducing reliance on histopathology.

Main Methods:

• Analysis of 234 LNs from 27 HNSCC patients, extracting 120 radiomic features per node.
• Systematic optimization of machine learning models, feature selection, and hyperparameters, including ensemble methods.
• Application of a Pareto front strategy for feature reduction and accuracy optimization.

Main Results:

• An optimal radiomics model achieved a balanced accuracy of 0.90, AUC of 0.93, and F1-score of 0.88.
• The high performance was obtained using only five selected radiomics features.
• The model was validated using 5-fold cross-validation.

Conclusions:

• The developed radiomics approach offers an interpretable and accurate method for noninvasive LN classification in HNSCC.
• This technique shows significant potential for integration into clinical radiology practice for improved HNSCC management.