Comparing Handcrafted Radiomics Versus Latent Deep Learning Features of Admission Head CT for Hemorrhagic Stroke...

Area of Science:

• Radiology
• Artificial Intelligence
• Medical Imaging Analysis

Background:

• Handcrafted radiomics extract predefined quantitative features from medical images.
• Deep neural networks learn features de novo through iterative training.
• Acute intracerebral hemorrhage (ICH) requires accurate prognostication for patient outcomes.

Purpose of the Study:

• To compare handcrafted radiomics and deep learning features for predicting 3-month outcomes and hematoma expansion in ICH.
• To evaluate the added value of deep learning features when combined with radiomics.
• To assess the potential of these methods for risk stratification in hemorrhagic stroke.

Main Methods:

• Utilized a multicenter cohort (n=866) for training/cross-validation and a single-center dataset (n=645) for external validation.
• Trained multiscale U-shaped networks for hematoma segmentation, extracting radiomics and two latent deep feature sets.
• Reduced features using Non-Negative Matrix Factorization (NMF) and applied six machine-learning classifiers.

Main Results:

• Combining latent deep features with radiomics numerically improved prediction performance for 3-month outcomes and hematoma expansion.
• Statistical significance in accuracy improvement was observed for predicting >3 mL hematoma expansion.
• Consistent, albeit modest, increases in prediction performance were noted across multiple classifiers.

Conclusions:

• Latent deep features extracted from head CT scans show potential for improving prognostication in ICH.
• The combination of radiomics and deep learning offers enhanced risk stratification at the individual level.
• Further research into deep learning for extracting clinically relevant information in hemorrhagic stroke is warranted.