Predictive performance of radiomic models based on features extracted from pretrained deep networks

Area of Science:

• Radiomics
• Medical Imaging
• Machine Learning

Background:

• Radiomic modeling traditionally uses generic texture and morphological features.
• Deep learning networks offer an alternative for extracting advanced features.
• Decisions in deep feature extraction methods can influence model outcomes.

Purpose of the Study:

• To assess the impact of various deep feature extraction choices on radiomic model predictive performance.
• To compare deep features against traditional generic features in radiomic modeling.

Main Methods:

• Trained radiomic models on ten public datasets using varied network architectures, feature extraction layers, slice selection, segmentation use, and aggregation methods.
• Employed a linear mixed model to quantify the influence of these choices.
• Compared deep feature models with generic feature models for predictive performance and feature correlation.

Main Results:

• Network architecture, feature extraction level, and using all slices significantly influenced model performance (p < 0.001).
• Segmentation use had a minor influence (p = 0.023), while aggregation method was insignificant (p = 0.774).
• Deep feature models did not significantly outperform generic feature models (p > 0.05); deep features showed moderate correlation (r=0.4) versus high correlation for generic features (r=0.89).

Conclusions:

• Model choices in deep feature extraction significantly affect predictive performance.
• Optimization of these choices via cross-validation is essential for maximizing model performance.
• Deep features offer diverse representations but do not inherently surpass generic features without careful selection.