MACHINE LEARNING APPLICATIONS IN RADIOLOGY: A SYSTEMATIC LITERATURE REVIEW

Abstract

Introduction of machine learning in radiology is a paradigm shift in diagnostic medicine, which has the potential to solve the growing imaging loads, deficit of staff, and differences in interpretive sensitivity. Nevertheless, the high rate of research development on the field requires a systematic synthesis to differentiate between the cumulative improvements and the paradigm shifts. This systematic review critically evaluates and summarizes the available empirical evidence on machine learning use in radiology regarding technical methodologies, clinical performance, strategies of workflow integration, and barriers to implementation across the larger imaging modalities. In accordance with PRISMA 2020, PubMed, Scopus, Web of Science, IEEE Xplore, and Cochrane Library were thoroughly searched and covered articles that were published within the timeframe of January 2019 to December 2025. Search search strategies were a combination of controlled vocabulary and free-text words based on machine learning, deep learning, radiomics, and diagnostic imaging. Research papers were incorporated in case they had original research on ML applications in radiology that presented valid performance measures. Quality of methodology was determined by QUADAS-2 and CLAIM checklists. The thematic analysis of a narrative synthesis was conducted because of methodological heterogeneity. Of the identified records (3,847), 187 studies were included. Thematic analysis showed that there were six key domains, including diagnostic accuracy comparison between ML and radiologists (n=42 studies), radiomics and quantitative imaging biomarkers (n=38), deep learning based on automated segmentation (n=35), multimodal integration with clinical data (n=24), workflow optimization and triage systems (n=29), and implementation science and human-AI interaction (n=19). Under the curve Pooled area under the curve values were between 0.82 and 0.96 between applications with much variations in study design, reference standards and methods of validation. External validation was only used in 23% of studies and prospective multicenters were few (12%). Machine learning shows great promise in increasing the accuracy of diagnoses, decreasing interpretation time, and obtaining quantitative imaging biomarkers that a human eye cannot perceive. Nevertheless, there is a translational gap between retrospective algorithm development and clinical application, which characterizes the field. Further studies are needed to focus on external validation, standardized reporting, prospective multicenter studies, and effective human-AI collaboration models in order to fulfill the potential of machine learning-enhanced radiology.