Image biomarkers and explainable AI: handcrafted features versus deep learned features.

Area of Science:

• Medical imaging analysis
• Artificial intelligence in healthcare
• Radiomics and deep learning

Background:

• Feature extraction and selection are crucial for radiomics and image biomarker discovery using architectures like convolutional neural networks (CNNs).
• Deep learning (DL) offers powerful feature extraction but requires careful consideration of dataset characteristics and feature types.

Purpose of the Study:

• To outline typical radiomics steps and CNN components for feature extraction.
• To discuss the impact of dataset size, diversity, and computational resources on model performance.
• To compare handcrafted features with deep learned features and highlight the importance of model explainability.

Main Methods:

• Description of radiomics workflows and CNN architectures (deep feature extraction vs. end-to-end).
• Discussion of dimensionality reduction techniques to address the curse of dimensionality.
• Exploration of dataset considerations, computational resources (GPUs), and federated learning.

Main Results:

• Deep learning approaches show outstanding performance but handcrafted features remain relevant.
• Dataset size and diversity significantly impact model stability and generalization, with representative datasets being essential.
• Non-DL methods offer superior explainability compared to DL, necessitating explainable AI (XAI) for complex tasks.

Conclusions:

• Key concepts for extracting reliable and robust image biomarkers from imaging features are provided.
• The study highlights differences between radiomics and representation learning, and the trade-offs between handcrafted and learned features.
• The clinical purpose of AI models must guide the selection of feature extraction methods.