Machine Learning in Radiomics: A Practical Guide for Modeling with Data-Driven Imaging Markers

• Activity Number: 551 - Statistical Machine Learning and Artificial Intelligence in Multi-Parametric Quantitative Imaging
• Type: Topic Contributed
• Date/Time: Thursday, August 11, 2022 : 10:30 AM to 12:20 PM
• Sponsor: Section on Statistical Computing
• Author(s): Xiaofeng Wang*
• Companies: Cleveland Clinic Lerner College of Medicine of Case Western Reserve University
• Keywords: Radiomics; Machine Learning; data-driven; prediction; imaging markers; guidelines

Abstract

Medical imaging parameters are used clinically to help detect disease, identify phenotypes, define longitudinal changes, predict outcomes, etc. We focus on derived features selected from radiomic analysis. Unlike studies in which features are evaluated for association with an outcome determined by an objective truth basis, in radiomic analysis a large number of features are extracted from radiographic medical images using data-driven algorithms. These radiomic features are referred to as data-driven imaging markers (DIMs) because they are quantitative measures discovered under a data-driven framework from images beyond visual recognition but evident as patterns irrespective of ground truth basis. Our aim is to set guidelines on how to build machine learning models using DIMs in radiomics and to apply and report them appropriately. We bring forth a list of recommendations, named RANDAM, for statistical analysis, modeling, and reporting in a radiomic study. We also provide simulation study results comparing variable selection methods and validation strategies along with a real case study in lung cancer research.