Healthcare’s environmental footprint is under increasing scrutiny, and imaging is a recognised contributor to sector-wide emissions. Artificial intelligence is now embedded across radiology workflows, supporting image analysis, reconstruction and reporting, which positions it as both an opportunity and a responsibility for sustainable practice. A scoping synthesis identified 13 publications on environmental sustainability in radiology AI published between 2018 and 2024. Reported impacts clustered around four themes: energy consumption, carbon footprint, computational resources and water use. Methods to measure and reduce these impacts are emerging, yet reporting remains uneven across the AI lifecycle. The evidence points to practical levers that can be implemented now, alongside clearer standards to improve transparency and comparability.
Evidence Base and Scope
The publications originated from the United States, China, Italy, the United Kingdom, Egypt, India and Portugal, with a gradual rise in output over time and most items appearing in the most recent year surveyed. The set comprised nine original research articles and four reviews. Modalities most frequently addressed included magnetic resonance imaging (MRI) and computed tomography (CT), with additional work touching on mammography and X-rays, several reviews did not specify modalities. Among the original research, deep learning dominated, with traditional machine learning represented in a smaller share. Reported tasks spanned segmentation, classification, diagnosis and image reconstruction. Toolchains typically relied on mainstream frameworks and accelerators.
Must Read: Hidden Environmental Footprint of Radiography and Fluoroscopy
Selection and extraction followed a scoping approach and incorporated comparative synthesis against prior reviews to situate the findings. Despite growing attention, environmental aspects were inconsistently documented beyond headline training runs. Data usage choices such as augmentation were noted without being framed systematically as sustainability interventions. Inference-time behaviour received limited coverage even though deployment scale can accumulate energy draw. No included source described structured end-of-life processes for models, datasets or pipelines, leaving a gap in lifecycle accounting. Only a minority specified impact-estimation tools, naming Carbon Tracker and the Green Algorithms calculator. Where methods were described, several counted only the final training pass, omitting hyperparameter searches and architecture exploration that materially contribute to energy use and emissions.
Metrics, Impacts and Gaps
Assessments grouped into energy consumption, carbon footprint, computational resource usage and water consumption. Energy was reported in kilowatt hours or joules, sometimes contextualised with power usage effectiveness (PUE) or equivalent car travel distance. Carbon footprint was expressed as greenhouse gas emissions in CO2-equivalent, with some translating these into notional tree sequestration capacity. Computational load was captured through training and runtime durations, epochs, floating-point operations, parameter counts and inference time. Water consumption was quantified in gallons, reflecting direct data centre cooling and the indirect demands of electricity generation.
Heterogeneity in architectures, datasets, hardware and measurement approaches precluded consistent numerical ranges. Several sources emphasised that headline figures typically exclude the iterative development work that precedes the reported training run, which underestimates real-world energy and carbon budgets. The inference phase was sparsely reported relative to training, despite the potential for cumulative energy use when workloads are scaled. Embodied carbon associated with hardware manufacturing, transport and disposal was identified as a blind spot, with limited estimates for production-phase emissions of GPUs and TPUs. Broader environmental impacts such as electronic waste, rare earth mineral extraction and land degradation were noted but remain underrepresented in model-centric calculators. Transparency on water use is uneven, and geographic variation in grid carbon intensity affects outcomes, underscoring the value of location-aware hosting and consolidated storage practices.
Practical Strategies and Governance
Actionable levers were reported across infrastructure, model design, training practice and deployment. Consolidating information technology infrastructure can reduce emissions by simplifying recycling and disposal and by lowering pollution associated with manufacturing. Cloud platforms may ease on-site hardware needs through economies of scale, while fog or edge computing reduces energy consumption, latency and bandwidth, which is advantageous for time-sensitive applications that process data close to acquisition. Extending hardware lifespan, transitioning to energy-efficient equipment or lightweight edge devices, and promoting responsible recycling were proposed to address embodied impacts.
On the algorithmic side, lightweight architectures with fewer parameters, compact convolutional networks, and optimisation techniques such as early stopping and efficient initialisation were linked to shorter training times and lower compute. Compression approaches, including quantisation and pruning, can reduce inference complexity without sacrificing task performance. Careful hyperparameter selection helps, though the search overhead itself should be reflected in reporting. Workload placement on fewer, highly utilised servers and edge scheduling tuned to regional demand were identified as additional efficiency measures.
Governance and transparency emerged as cross-cutting needs. Authors recommended integrating environmental indicators alongside accuracy and other performance metrics in AI reporting. Proposals included a radiology-specific ecolabel to standardise disclosures on energy and carbon and inclusion of core indicators—such as CO2-equivalent emissions, parameter counts, training and inference times, PUE and applied efficiency techniques—within model documentation. Collaboration across institutions can limit redundant training and storage and support hosting energy-intensive tasks where cleaner energy is available. Education for radiology professionals on green AI principles was highlighted to embed efficiency throughout development and deployment.
Evidence on the environmental footprint of radiology AI is expanding, with clear priorities for action. Lightweight models, efficient training routines, and thoughtful deployment across cloud, fog and edge can reduce energy use and emissions, while acknowledging water consumption and embodied impacts. Current reporting practices often omit development iterations, inference at scale and end-of-life, limiting comparability and masking total costs. Embedding environmental indicators into routine evaluation and documentation, supported by standardised disclosure and collaborative infrastructure choices, can help radiology teams align AI-enabled imaging with measurable progress toward lower-impact practice.
Source: European Journal of Radiology
Image Credit: iStock
