Background
The healthcare sector is experiencing significant technological transformation, particularly in medical imaging (Doo et al., 2024). Over the past four years, cloud migration and Artificial Intelligence (AI) adoption have surged, with 219 EU CE-marked AI solutions for radiology available as of 2024 (Health AI Register,2024). However, while these advancements offer clinical benefits, they also raise environmental and economic considerations (Doo et al., 2024; Picano et al.,2022). Depending on their complexity, AI solutions can require significant energy, increasing the carbon footprint of medical imaging workflows (Truhn et al.,2024).
Many healthcare institutions are now committing to carbon neutrality, acknowledging the environmental impact of high-energy medical imaging processes. However, balancing technological progress with sustainability remains a challenge. This study evaluates the carbon footprint of a modern CT imaging workflow, comparing cloud-based and on-premises Picture Archiving and Communication System (PACS) and AI diagnostic solutions, while considering cost implications and cybersecurity risks.
Methodology
This study quantifies the carbon footprint of a CT imaging workflow across three components:
- CT imaging: Energy use, consumables, and machine manufacturing emissions.
- PACS processing: Cloud-hosted vs. on-premises.
- AI diagnostic solutions: Cloud-hosted vs. on-premises.
Data was collected from a global hyperscale cloud service provider (CSP), a large Dutch hospital (1,500 beds), and a European AI diagnostic solutions company. This analysis also incorporates independent perspectives from academic researchers and industry analysts to enhance neutrality. Hospitals that have opted against cloud migration were interviewed to understand their rationale.
Results
A CT imaging workflow using an AI diagnostic solution (cloud-hosted) emits 11.32 kgCO2e per examination. In a 1,500-bed hospital conducting 10,200 CT scans annually per machine, annual emissions total 115,435.31 kgCO2e per machine.
Cloud-hosted PACS reduces emissions by 252.33 kgCO2e annually per machine. Table IV illustrates emission differences
Table IV : PACS Hosting Emissions Comparison.
AI diagnostic solutions also benefit from cloud hosting, eliminating 200 kgCO2e per year compared to on-premises AI solutions. However, hospitals retaining on-premises AI solutions cite concerns over data sovereignty, regulatory compliance, and cybersecurity risks.
Risks and Alternatives
While cloud adoption reduces emissions, potential drawbacks exist:
Cybersecurity vulnerabilities: Data breaches remain a top concern, with healthcare cyberattacks increasing 38% in 2023.
Data sovereignty: Some hospitals prefer on-premises PACS to ensure compliance with national regulations.
Long-term costs: Cloud services appear cost-efficient in the short term but require continuous subscription fees, which can accumulate significantly over time.
Energy use by hyperscale data centres: Though many operate on renewable energy, their total energy demand is still substantial.
Alternative strategies for institutions hesitant about cloud adoption include hybrid cloud models, renewable-powered on-premises data centres, and energy-efficient GPU upgrades for AI workloads.
Cost Comparison
A total cost of ownership (TCO) analysis was conducted, comparing cloud-based and on-premises solutions over five years. Independent financial assessments revealed:
Cloud services can be 30% cheaper if optimised, but if not carefully managed, their costs can exceed those of on-premises solutions.
Depending on the processing requirements, AI workload hosting can be more cost-effective on CPUs rather than GPUs.
Hyperscale cloud providers offer higher Power Usage Efficiency (PUE) (1.22) compared to enterprise data centres (1.84), increasing energy efficiency but requiring vendor lock-in.
Hospitals should obtain assessed quotations, optimise service plans, and consider long-term scalability to make informed financial decisions.
Recommendations
1. Transparency in Funding and Vendor Affiliations
Healthcare institutions must disclose affiliations with cloud providers and AI vendors to enhance credibility, such as Microsoft Azure, Google Cloud, IBM Watson Health, Siemens Healthineers Teamplay Cloud, Green AI Cloud and AWS. Independent oversight bodies should verify cost assessments and performance claims.
2. Define "Cost" at an Organisational Level
Environmental impact should be integrated into cost assessments. This requires leadership buy-in, clear communication, and structured implementation at all levels.
3. Tailor Technological Adoption Strategies
Rather than a one-size-fits-all approach, hospitals should assess:
Workload demands and infrastructure capacity.
AI adoption timelines based on grants and funding.
Energy-efficient alternatives before cloud migration.
4. Vendor Diversification
Expanding beyond major hyperscale CSPs, this study includes additional vendors offering regional cloud hosting and energy-efficient AI solutions:
Scaleway: Provides European cloud hosting with strict data sovereignty compliance.
Green AI Cloud: Focuses on carbon-neutral AI processing for healthcare applications.
Siemens Healthineers Teamplay Cloud: A healthcare-dedicated cloud service offering hybrid models.
5. Cybersecurity and Data Governance Frameworks
Hospitals should implement robust security policies, including:
End-to-end encryption.
Compliance with regional data protection laws.
Regular security audits to assess vulnerabilities.
Conclusion
Medical imaging is at a crossroads where sustainability, technological innovation, and economic considerations must be carefully balanced. While cloud adoption can significantly reduce emissions, cybersecurity, cost, and regulatory concerns remain critical. Hospitals should adopt a strategic, case-by-case approach to technology integration, ensuring that financial, environmental, and operational needs are met.
By diversifying vendors, refining cost assessments, and adopting hybrid models, healthcare institutions can make sustainable and financially responsible decisions in medical imaging workflows.
Title Image Credit: Freepik
References:
Doo, F. X., Vosshenrich, J., Cook, T. S., Moy, L., Almeida, E. P. R. P., Woolen, S. A., Gichoya, J. W., Heye, T., & Hanneman, K. (2024). Environmental Sustainability and AI in Radiology: A Double-Edged Sword. Radiology, 310(2), e232030. https://doi.org/10.1148/radiol.232030
Health AI Register. (2024). Products. Health AI Register. https://radiology.healthairegister.com/products/?subspeciality=All&modality=All&ce_under=All&ce_class=All&fda_class=All&sort_by=ce%20certification&search=
Picano, E., Mangia, C., & D’Andrea, A. (2022). Climate Change, Carbon Dioxide Emissions, and Medical Imaging Contribution. Journal of Clinical Medicine, 12(1). https://doi.org/10.3390/jcm12010215
Truhn, D., Müller-Franzes, G., & Kather, J. N. (2024). The ecological footprint of medical AI. European Radiology, 34(2), 1176–1178. https://doi.org/10.1007/s00330-023-10123-2
