Image-Level Data Augmentation for Radiomics-Based Classification of Vital Versus Non-Vital Persistent Cervical Lymph...

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Oncology

Background:

• Distinguishing vital from non-vital persistent cervical lymph nodes after chemoradiotherapy in head and neck squamous cell carcinoma (HNSCC) is clinically challenging.
• Radiomics analysis of CT scans offers potential for objective assessment.
• Improving the accuracy of radiomics classification is crucial for treatment guidance.

Purpose of the Study:

• To investigate the impact of image-level data augmentation on CT-based radiomics classification of persistent cervical lymph nodes in HNSCC.
• To optimize augmentation strategies and radiomics pipeline parameters for improved classification performance.

Main Methods:

• Evaluated eight data augmentation strategies and their combinations in 55 HNSCC patients.
• Employed Bayesian hyperparameter tuning with Optuna for parameter optimization.
• Assessed a radiomics pipeline including feature extraction, selection, and classification using 5-fold cross-validation, ranking configurations by a composite score (mean of AUC, ACC, and F1-score).

Main Results:

• Feature selection alone improved the composite score from 0.659 to 0.742.
• The best augmented configuration (Window Contrast Variation) achieved a composite score of 0.803 and an AUC of 0.831.
• The combined approach of feature selection and optimized augmentation yielded significant gains over baseline and feature selection alone.

Conclusions:

• Feature selection combined with optimized image data augmentation shows promise for enhancing radiomics-based classification of cervical lymph nodes in HNSCC.
• While promising, individual augmentation strategy differences require validation in larger patient cohorts to achieve statistical significance.