Comparative Evaluation of Radiomics and Deep Learning Models for Disease Detection in Chest Radiography.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• AI is transforming medical imaging diagnostics.
• Chest radiography analysis for COVID-19, lung opacity, and viral pneumonia benefits from AI.
• Radiomics and deep learning (DL) offer distinct approaches to AI in medical imaging.

Purpose of the Study:

• To comprehensively evaluate and compare the diagnostic performance of radiomics-based and deep learning-based AI models for chest radiography.
• To assess the impact of sample size on the performance of different AI models.
• To provide data-driven recommendations for AI model selection in clinical diagnostic settings.

Main Methods:

• Systematic comparison of radiomics models (Decision Trees, Gradient Boosting, Random Forests, SVMs, MLPs) and DL models (InceptionV3, EfficientNetL, ConvNeXtXLarge).
• Evaluation of diagnostic performance across varying sample sizes (e.g., 24 and 4000 samples).
• Statistical analysis using Scheirer-Ray-Hare tests and Mann-Whitney U tests with Bonferroni correction.

Main Results:

• Deep learning models, such as EfficientNetL and InceptionV3, demonstrated superior performance (higher AUC) compared to radiomics models, particularly with larger sample sizes.
• EfficientNetL achieved an AUC of 0.839 at 24 samples, outperforming SVM (0.762).
• InceptionV3 reached an AUC of 0.996 at 4000 samples, significantly higher than Random Forest (0.885).
• Statistical tests confirmed significant effects of model type and sample size on performance metrics, favoring DL models.

Conclusions:

• Deep learning models offer superior diagnostic performance and scalability in medical imaging AI, especially as data availability increases.
• Radiomics-based models retain utility in low-data scenarios.
• The study provides statistically validated guidance for selecting appropriate AI models in diverse clinical environments for chest radiography analysis.