Machine Learning Bias in Medical Imaging: A Scoping Review

Background and Aims: Machine learning, which is a subset of artificial intelligence, is increasingly being utilized in numerous different areas, and particularly in the medical profession. However, its usage comes with a potential for malfunction and bias. The present study examines errors and biases in machine learning in the case of medical imaging. Methods: This scoping review was conducted based on the framework proposed by (Arksey & O'Malley, ۲۰۰۵) To identify relevant studies, the following keywords were used: Bias, Error, Machine Learning Bias, Algorithm Bias, machine learning, ML, Artificial intelligence, AI, medical imaging, radiology, imaging, radiography, Diagnostic Imaging, Diagnosis, Magnetic Resonance Imaging, MRI, Computed Tomography, CT, CT scan, Ultrasonic, X-ray. Searches were performed in PubMed/MEDLINE, EMBASE, Scopus, Web of Science, and IEEE Xplore. Following the examination of titles and abstracts, ۴۵ articles were incorporated in the research. Results: Bias in the input datasets, the application of training data with constrained or non-standard samples, disparities in racial and demographic attributes (such as age, gender, socioeconomic standing, lifestyle, and genetic background), fluctuations in image quality, luminance, and resolution (MRI, CT, X-ray), as well as differences between imaging devices, were among the most significant sources of bias in machine learning-based algorithms in the reviewed studies. Conclusion: Executing requisite strategies to alleviate these biases is crucial not solely for enhancing the precision of medical diagnosis and analysis but also for guaranteeing the just provision of healthcare services.

کلیدواژه ها:

Bias ، Machine Learning ، Artificial Intelligence