Causal Inference in Radiomics: Framework, Mechanisms, and Algorithms.

Area of Science:

• Radiomics
• Machine Learning
• Causal Inference

Background:

• Machine learning in radiomics offers flexible prognostic models but suffers from a "black-box" nature, hindering mechanistic understanding.
• Current radiomics approaches lack methods for estimating causal effects from latent exposures.

Purpose of the Study:

• To develop an inferential framework for estimating causal effects using radiomics data.
• To address the challenge of latent exposures in radiomics analysis.
• To enhance mechanistic and phenomenological understanding of clinical outcomes predicted by radiomics.

Main Methods:

• Development of a novel inferential framework tailored for radiomics data.
• Leveraging a multivariate version of partial least squares (PLS) for causal effect estimation.
• Application of the framework to radiomics datasets from osteosarcoma and glioblastoma.

Main Results:

• The proposed framework enables causal effect estimation from radiomics data, even with latent exposures.
• Demonstrated the utility of multivariate PLS for causal inference in this context.
• Successfully applied the methodology to real-world cancer datasets.

Conclusions:

• The developed framework offers a pathway to move beyond predictive radiomics towards causal understanding.
• This approach can improve mechanistic insights into prognostic factors identified through radiomics.
• The methodology shows promise for applications in various cancer types, including osteosarcoma and glioblastoma.