RadiomiX for Radiomics Analysis: Automated Approaches to Overcome Challenges in Replicability.

Area of Science:

• Medical Imaging Analysis
• Machine Learning in Healthcare
• Radiomics Research

Background:

• Radiomic analysis involves complex model selection, impacting result reliability.
• Automated methods are needed to streamline and validate radiomic model performance.
• Current radiomic models face challenges in generalizability and consistent performance.

Purpose of the Study:

• To develop and validate RadiomiX, an algorithm for automated radiomic model selection and validation.
• To enhance the reliability and performance of radiomic models through systematic testing.
• To simplify the decision-making process in radiomics.

Main Methods:

• RadiomiX systematically evaluates combinations of classifiers and feature selection methods.
• The framework utilizes multiple train-test splits and K-fold cross-validation for robust assessment.
• Validation was performed on four public datasets (lung nodules, breast cancer, hepatic encephalopathy) using CT, PET/CT, and MRI.

Main Results:

• RadiomiX demonstrated superior performance across all four datasets, achieving high AUC and accuracy.
• The algorithm significantly outperformed original published models (p < 0.001 for LLN/SLN).
• Re-evaluation of original models using cross-validation showed substantial performance decrease, highlighting RadiomiX's advantages.

Conclusions:

• Systematic model combination testing with RadiomiX significantly improves radiomic model performance.
• Automated machine learning offers a pathway to more reliable and higher-performing radiomic models.
• RadiomiX enhances the robustness and generalizability of radiomic findings.