Deep Features from Pretrained Networks Do Not Outperform Hand-Crafted Features in Radiomics

Area of Science:

• Radiomics and Medical Imaging Analysis
• Machine Learning in Healthcare
• Computational Pathology

Background:

• Deep learning models can extract complex features from medical images.
• Hand-crafted features have traditionally been used in radiomics.
• The potential of deep features from pretrained networks in radiomics remains an active area of research.

Purpose of the Study:

• To compare the predictive performance of radiomics models using deep features versus hand-crafted features.
• To evaluate the impact of combining deep and hand-crafted features on model performance.
• To assess the influence of pretraining data (medical vs. ImageNet) on deep feature performance.

Main Methods:

• Models were trained on ten radiological datasets.
• Five feature selection methods and three classifiers were employed.
• Performance was evaluated using Area Under the ROC Curve (AUC) and other metrics.

Main Results:

• Deep features did not yield significantly higher AUC than hand-crafted features (0.775 vs. 0.789).
• Combining morphological features with deep features improved overall model performance (AUC +0.02).
• Incorporating all hand-crafted features with deep features further improved overall performance (AUC +0.034), but not the best model's performance.

Conclusions:

• Radiomics models using deep features from pretrained networks do not inherently outperform models using hand-crafted features.
• Combining feature types can enhance overall predictive performance, though benefits may vary for top-performing models.
• Pretraining data source (medical vs. ImageNet) did not confer a significant advantage for deep feature extraction in this radiomics context.