Exploring Environmental Sustainability of Artificial Intelligence in Radiology: A Scoping Review

Synthesizing Evidence on Energy, Carbon, and Water Footprints of AI in Medical Imaging

M. Champendal; B. Lokaj; Valentin Durand de Gevigney; G. Brulé; Jamil Zaghir; P. Boiko; Christian Lovis; H. Müller; J. Schmid; R. Ribeiro· European Journal of Radiology· 2025· DOI 10.1016/j.ejrad.2025.112558

The core problem

Artificial intelligence (AI) is increasingly integrated into radiology, enhancing diagnostic accuracy and workflow efficiency. However, the environmental implications of training and deploying AI models—particularly deep learning (DL) and large language models (LLMs)—have not been sufficiently studied. This scoping review aims to synthesize existing literature on the environmental sustainability of AI in radiology and highlight strategies proposed to mitigate its impact. The review addresses a critical gap: while AI's clinical benefits are well-documented, its carbon footprint, energy consumption, and water usage remain underexplored. The objective is to map the current evidence, identify key metrics, and propose actionable recommendations for sustainable AI development in medical imaging.

Innovation

The 13 included studies revealed four major themes: energy consumption (n=10), carbon footprint (n=6), computational resources (n=9), and water consumption (n=2). Reported metrics included CO2-equivalent emissions, training time, power use effectiveness (PUE), equivalent distance travelled by car, energy demands, and water consumption. For instance, training a single DL model can emit up to 284 kg CO2eq, equivalent to 315 km driven by an average car. Energy consumption is often measured in kilowatt-hours (kWh), with some studies reporting training times exceeding 100 hours on GPUs. Carbon footprint calculations frequently use the formula: C O_{2} e q = E \times C I where E is energy consumption (kWh) and C I is the carbon intensity of the energy source (kg CO2eq/kWh). Computational resources are quantified via floating-point operations (FLOPs) or GPU hours. Water consumption, though less studied, is estimated using water usage effectiveness (WUE) metrics, with data centers consuming millions of liters for cooling. The review also identified strategies to enhance sustainability: lightweight model architectures (e.g., MobileNet, EfficientNet), quantization and pruning, efficient o

Artificial intelligence (AI) is increasingly integrated into radiology, enhancing diagnostic accuracy and workflow efficiency. However, the environmental implications of training and deploying AI models—particularly deep learning (DL) and large language models (LLMs)—have not been sufficiently studied. This scoping review aims to synthesize existing literature on the environmental sustainability of AI in radiology and highlight strategies proposed to mitigate its impact. The review addresses a critical gap: while AI's clinical benefits are well-documented, its carbon footprint, energy consumption, and water usage remain underexplored. The objective is to map the current evidence, identify key metrics, and propose actionable recommendations for sustainable AI development in medical imaging.

The review followed the Joanna Briggs Institute (JBI) methodology for scoping reviews. A systematic search was conducted across MEDLINE, Embase, CINAHL, and Web of Science, focusing on English and French publications from 2014 to 2024. Search terms targeted AI, environmental sustainability, and medical imaging. Eligible studies addressed the environmental sustainability of AI in medical imaging. Conference abstracts, non-radiological or non-human studies, and unavailable full texts were excluded. Two independent reviewers screened titles, abstracts, and full texts, while four reviewers performed data extraction and analysis. The search identified 3,723 results, of which 13 met inclusion criteria: nine research articles and four reviews. Data were charted according to themes: energy consumption, carbon footprint, computational resources, and water consumption. The review also extracted reported metrics and proposed mitigation strategies. The PRISMA-ScR flow diagram (Figure 1) summarizes the selection process.

Why it matters

The findings indicate that research on sustainable AI in radiology remains scarce but is rapidly growing. The predominance of energy and carbon metrics suggests a focus on operational efficiency, while water consumption is largely overlooked. The proposed strategies—lightweight architectures, quantization, and pruning—can reduce energy use by up to 80% without significant performance loss. However, trade-offs exist: quantization may slightly degrade accuracy, and cloud computing's sustainability depends on the provider's energy mix. The lack of standardized reporting hinders comparability across studies. For example, some studies report CO2eq per inference, others per training run, and few include life cycle assessment (LCA). The review highlights the need for transparent, consistent metrics and an eco-label to guide radiologists and developers. Future research should explore the environmental impact of LLMs and vision transformers (ViTs) in radiology, as these models are more resource-intensive. Additionally, policies incentivizing green AI—such as carbon-aware scheduling and renewable energy procurement—are essential. The review's limitations include the exclusion of non-English/French studies and conference abstracts, which may contain relevant data. Despite these limitations, the synthesis provides a foundation for integrating environmental sustainability into AI development in radiology.

Who should read this

CS practitioners and researchers

Opening member content…