Are deep models in radiomics performing better than generic models? A systematic review

Area of Science:

• Radiomics and Medical Imaging Analysis
• Artificial Intelligence in Healthcare
• Machine Learning for Diagnostic Imaging

Background:

• Radiomics traditionally uses generic models (GMs) with predefined features.
• Deep learning models (DMs) have emerged as a recent advancement in radiomics.
• The comparative performance of DMs versus GMs in radiomics remains unclear.

Purpose of the Study:

• To systematically compare the performance of deep learning models (DMs) against generic models (GMs) in radiomics.
• To evaluate differences in diagnostic accuracy, measured by the area under the curve (AUC).
• To analyze the impact of validation strategies (internal vs. external) on model performance.

Main Methods:

• A comprehensive literature search was conducted on PubMed and Embase.
• Included studies published between 2017 and 2021 focusing on radiomics.
• Performance metrics, primarily AUC, were compared between DMs and GMs.

Main Results:

• Deep learning models (DMs) outperformed generic models (GMs) in 74% of internal and 65% of external validation studies.
• Fused models demonstrated superior performance in 72% of internal and 63% of external validations.
• A notable percentage of studies (26%) showed DMs not outperforming GMs, highlighting variability.

Conclusions:

• Deep learning models generally show superior performance over generic models in radiomics applications.
• The effectiveness of DMs varies, with GMs still outperforming them in a subset of cases.
• Further research is needed to optimize DM performance and understand limitations in radiomics.