A Novel Feature Engineering Method Based on Latent Representation Learning for Radiomics: Application in NSCLC...

Area of Science:

• Medical Imaging
• Machine Learning
• Computational Biology

Background:

• Radiomics extracts quantitative features from medical images for clinical outcome prediction.
• Current feature engineering methods struggle to effectively use the heterogeneity of radiomics features.

Purpose of the Study:

• To introduce latent representation learning as a novel feature engineering approach for radiomics.
• To improve the utilization of feature heterogeneity in radiomics.

Main Methods:

• Developed a latent representation learning method to reconstruct latent space features from original shape, intensity, and texture features.
• Utilized a hybrid loss function (clustering-like and reconstruction loss) to optimize latent space features.
• Validated the method on a multi-center non-small cell lung cancer (NSCLC) subtype classification dataset.

Main Results:

• Latent representation learning significantly improved classification performance compared to traditional methods (PCA, Lasso, etc.) on an independent test set (p<0.001).
• Demonstrated significant improvements in generalization performance on two additional test sets.
• Showcased superior performance across various machine learning classifiers.

Conclusions:

• Latent representation learning is a more effective feature engineering method for radiomics.
• This approach has the potential to be a general technology applicable to a wide range of radiomics research.
• The method effectively addresses the limitations of current feature engineering techniques.