You are here: Machine Learning for Precision Health»Radiogenomics: Mining the Correlation Between Radiological Imaging Features and Gene Expression

Radiogenomics: Mining the Correlation Between Radiological Imaging Features and Gene Expression

For diseases and disorders with genetic components and salient imaging phenotypes, the joint analysis of the data holds potential to improve disease understanding.However, such datasets have an overwhelming number of features and feature combinations, making an exhaustive study impractical.In my research I address this need by developing advanced algorithms that rank features by their importance for a given disease outcome and that identify a sparse set of features that can be used to train highly accurate predictors of individualized outcomes.In recent work I have embedded one of the most powerful regressors, the decision forest, in a model reduction framework endowing it with sparsity to reduce model complexity and reduced variance.

Radiogenomics Diagram

Radiogenomics for improved disease understanding. We extract exquisite radiological image features, derive genomics features such as single nucleotide polymorphisms (SNPs) from a genome wide association study (GWAS) and then build machine learning based predictions of long term patient outcome, such as conversion from MCI to Alzheimer’s in 18mo. The features used by the predictor can then be examined for potential gene targets and imaging based biomarkers.

Applying it to Alzheimer’s, the approach automatically and simultaneously selects (Figure) imaged brain regions and SNP variants from a GWAS panel that agree with clinical literature and predicts cognitive outcome scores with high accuracy.

Our current work addresses: increasing feature selection accuracy, quantifying feature interaction, and testing the improvements on clinical applications, including schizophrenia and ASD.