Machine Learning Reconstruction: What You Need to Know

Machine learning algorithms, which learn from data rather than explicit programming, have been increasingly applied in radiology due to the availability of large datasets and parallel computational power. Their use in image reconstruction offers the potential to improve image quality, reduce artifacts, decrease acquisition times for MRI, and lower radiation doses for CT beyond what would be achievable with standard reconstruction approaches. However, these approaches have limitations including instability; the risk of introducing new artifacts; dependence on the quantity, quality, and diversity of the training data; and lack of explainability. Although machine learning can improve the efficiency and quality of imaging, improvements may not lead to better patient outcomes, and many commercially available machine learning–based products require further validation. The potential of machine learning image reconstruction should be harnessed with care, and the radiology community must remain actively engaged in guiding its development and implementation.