Improving radiomic model reliability using robust features from perturbations for head-and-neck carcinoma.

Area of Science:

• Radiomics
• Medical Imaging Analysis
• Machine Learning in Oncology

Background:

• Radiomic models are increasingly used, but the impact of feature robustness is not fully understood.
• High-robust radiomic features are recommended for modeling, yet their effect on model performance requires evaluation.

Purpose of the Study:

• To evaluate the impact of screening low-robust radiomic features on model robustness and generalizability.
• To validate findings across multiple datasets and clinical tasks in head and neck cancer.

Main Methods:

• Utilized computed tomography images from 1,419 head-and-neck cancer patients across four datasets.
• Quantified radiomic feature robustness using intra-class correlation coefficient (ICC) via a perturbation method.
• Constructed radiomic models using all features, good-robust (ICC > 0.75), and excellent-robust (ICC > 0.95) features, employing filter-based selection and Ridge classification.

Main Results:

• Model robustness (ICC) significantly improved from 0.65 to 0.78 (good-robust) and 0.91 (excellent-robust) (P<0.0001).
• Model generalizability improved, with reduced train-test AUC differences from 0.21 to 0.18 (good-robust) and 0.12 (excellent-robust) (P<0.001).
• Good-robust features achieved the best average AUC (0.58) on unseen datasets.

Conclusions:

• Incorporating robust radiomic features significantly enhances model robustness and generalizability.
• Model robustness requires verification even when using robust features.
• Overly strict feature selection may hinder optimal model performance by reducing discrimination power.