ARRS 2022 Abstracts

RETURN TO ABSTRACT LISTING


E1168. Basic Steps in Machine Learning: A Primer for Current and Future Radiologists
Authors
  1. Connie Ju; University of California at Los Angeles
  2. Tracie Kong; University of California at Los Angeles
  3. Nancy Pham; University of California at Los Angeles
Background
The increased focus on machine learning (ML) applications within the radiology workflow has promoted the introduction of basic ML concepts into the radiology curriculum. As targeted users of ML products, radiologists should maintain an introductory understanding of project development and implementation, yet there are few resources to date that provide basic concepts to those with little or no background in data science or programming. Understanding technical limitations of these products begins with fundamental understanding of the ML pipeline, thus learning basic methodology and terminology will help position radiologists and radiology trainees to critically appraise new research developments and products in the field. In this exhibit, we review the basic steps to ML in radiology including identifying use case scenarios, data collection and preprocessing, model training, and workflow integration. We will build upon this knowledge by elucidating the limitations that arise along each step of the pathway and how they affect the final ML product.

Educational Goals / Teaching Points
The goals of this exhibit are to introduce commonly encountered terminology used within the ML pipeline, including the differences between artificial intelligence, ML, and deep learning; outline basic steps within the ML pipeline (i.e., identifying the ML task, data collection and preprocessing, model development, model training, performance evaluation, and workflow integration); and understand how technical limitations in the ML workflow contribute to pitfalls of ML products.

Key Anatomic/Physiologic Issues and Imaging Findings/Techniques
In this exhibit, we review the basic steps for implementation of a ML project at an introductory level that is accessible for all radiologists and radiology trainees. We highlight potential use case scenarios to illustrate how implementation of ML projects can help guide clinical decision-making. Lastly, we demonstrate how limitations of the ML workflow translate to the current challenges facing the field. This includes the processing of large imaging data sets, reproducibility, assessment of radiogenomic associations, and clinical translation.

Conclusion
ML has established a new frontier in radiology that can impact the direction of radiology research and clinical workflow in the years to come. It is beneficial for practicing radiologists and radiology trainees to familiarize themselves with the basic principles of ML and commonly encountered terminology in order to critically appraise future ML applications in image analysis.