Two Sides of a Coin: AI and Sustainability in Radiology

Medical imaging is a significant consumer of energy, a fact that becomes startlingly clear when you learn it accounts for about 1% of greenhouse gas (GHG) emissions. The primary culprits have been the production and operation of MRI, CT, and other similar equipment. Now, AI's rapidly increasing energy demand is joining the ranks. We're at a crossroads of innovation and climate change, a classic double-edged sword. A compelling examination of this dilemma is offered in Florence Doo et al.’s publication at the University of Maryland.

Let's look at the downsides first.

Unsaturated Hunger: AI's High Energy Consumption and GHG Emissions

AI's energy consumption hinges on two main factors: computational efforts and data storage capacity needed for training and deploying AI models. To give you an idea, training a new AI model can emit as much GHG as the lifetime emissions of 5 cars, including their manufacture. Today's data storage systems demand more energy than the entire airline industry, making a significant impact on global climate change.

Not just an environmental problem: AI's rise is causing trouble for the power grid

This technological revolution is causing immediate, tangible issues: our power grid is being pushed to its limits. In the US, Georgia is already predicting a 17-fold increase in energy consumption. Consider Oracle's recent announcement of an AI data center the size of eight Boeing 747s. These colossal data centers need power, and increasing gas-fired power plants to meet this demand is hardly a viable solution.

The Battle is Not Lost: Smarter Approaches to Tackle the Problem

Despite these challenges, there are several innovative strategies in the works:

• Reduce the computational power used for AI training and deployment, finding a balance between speed and energy use.
• Employ tiny machine learning, which is a more energy-efficient approach.
• Reuse open-source models for AI training to save on energy.
• Avoid developing redundant AI solutions; why have twelve AIs for the same purpose?
• Implement tiered data storage, using quick and energy-intensive devices only for frequently accessed data.
• Utilize data compression techniques and eliminate redundant data.
• Consider cooler climates for data centers to reduce cooling energy demands.

Moving to the Upside: AI's Potential to Save Energy and Reduce Emissions

AI also has a promising role in energy conservation and reducing GHG emissions

1. Optimizing MRI and CT machine operations is essential for energy conservation. Surprisingly, about one-third of MRI energy consumption occurs when it's in an off state; for CT, it's two-thirds. AI tools can be strategically utilized to forecast when these machines are idle and determine the most efficient times to shut them down and restart them, significantly impacting energy savings.
2. Reducing image acquisition time is another critical area. As scan duration directly correlates with energy usage, employing AI for techniques like de-noising can hasten the process. AI can also automate the planning of imaging planes, effectively trimming down both the time and energy required for scans. An added benefit of these AI applications is enhancing patient comfort by reducing their time in the machine.
3. Enhancing the utility of low-field MRI is another promising aspect. By using AI tools, low-field MRI can evolve into a low-energy option compared to conventional MRI. This not only saves energy but could also help expand MRI access globally, thus potentially improving worldwide health outcomes.
4. Synthetic CT generated from MRI data presents an opportunity to reduce the number of CT scans needed, particularly in procedures like cancer staging. This approach could lead to a reduction in costs, energy usage, and increase patient comfort by minimizing the number of scans required.
5. Another area where AI can make a significant impact is in decreasing low-value imaging. Presently, about 20% of imaging procedures offer little to no value for patient diagnosis or treatment. AI tools can analyze extensive data sets - including past imaging, treatments, blood tests, and vital parameters - to advise on future imaging decisions. This would be invaluable assistance for both radiologists and referring physicians.
6. AI can also aid in reducing travel-related GHG emissions through optimal scheduling. By effectively managing and synchronizing doctor appointments and other tests with radiology scans, AI can streamline the entire process, making healthcare more efficient and environmentally friendly.

What about contrast agents (CA)?

Tough news for all die hard contrast agent fans (so basically all radiologists in their right mind). Looking at the data the global use of CA is steadily increasing. Each year, around 300 million CT scans are performed worldwide, with about 40% requiring contrast enhancement, averaging 100 ml per scan. This amounts to a staggering 12 million liters of CA used globally every year. Specifically, in Germany, about 70 tons of CAs are excreted into the River Rhine annually.

Sounds bad, but what's actually the link between CA situation and sustainability?

• Water Pollution: One of the major issues is the water pollution caused by patient excretion post-scan.
• Greenhouse Gas Emissions: The production, packaging, and delivery of CAs contribute to GHG emissions. Moreover, certain CAs, like sulfur hexafluoride, are potent greenhouse gases themselves—22 times more so than CO2.

While the necessity of CAs for diagnostic accuracy in many cases is undeniable, AI tools offer innovative solutions to mitigate these issues:

1. Optimizing Image Contrast: AI can enhance image contrast or employ virtual contrast techniques, potentially eliminating the need for physical CAs.
2. Clinical Decision Support: AI can aid clinicians in determining which imaging studies might not require contrast, thereby reducing CA usage.
3. Low-dose contrast approaches via restoring signal-to-noise ratio: AI-enhanced techniques can lower the required volume of CAs.

⛔️Obstacles in Implementation:

Implementing these AI solutions for radiology sounds great in theory, but the reality of putting these ideas into practice brings its own set of challenges:

1. Lack of GHG Data for AI in Radiology: Right now, we don’t have enough specific data on how much greenhouse gases are emitted by AI applications in radiology. Without solid data, it's tough to make solid decisions.
2. No Standard Measurements: We’re also missing standardized ways to measure the environmental impact of AI software. This makes it hard to set any kind of guidelines for choosing eco-friendly AI options.
3. Too Many Similar AI Tools: The market’s flooded with too many similar AI models and datasets. This redundancy isn't helping anyone and just adds to the problem.
4. Big Upfront Costs: Getting into the AI game, especially with sustainability in mind, means everyone involved has to put up a lot of money upfront.

🧠Overcoming These Obstacles:

Fortunately, the answers to these problems are already there, and to be fair, they've often been the solutions to humanity's (r-)evolutionary challenges: collect data, share information, and share resources. So, how do we do this?

• Gather and Standardize Data: We need to start by collecting precise data on AI's GHG emissions in radiology and come up with a standard way to measure this.
• Set Guidelines for Sustainable AI: Once we have the right measurements, we can make guidelines to help pick the most eco-friendly AI software and infrastructure.
• Share Resources: Sharing stuff like code, datasets, and learning methods could help spread out the cost of investing in AI, right from its development to when it’s actually used.
• Incentives for Eco-Friendly AI: We could encourage the development of sustainable AI by setting up incentives. This could be things like needing to prove an AI tool is sustainable before it gets published in a scientific paper or tweaking rules to make non-sustainable AI less attractive on the market.

✌️My Personal Opinion:

AI in radiology is here to stay, and we need to steer it towards a sustainable future. To me, the keys, as stated in the publication, are sharing and collaborating. It's a clear win-win because this approach can be the most cost-effective and sustainable. Ultimately, this is a matter of framing, attitude, and communication among stakeholders. Continuous data collection is another crucial point; we need to establish benchmarks for the sustainability of AI in radiology. This will help us set the right incentives and solidify sustainability as a key factor in choosing the most suitable AI software.

Sources:

1️⃣ Doo FX, Vosshenrich J, Cook TS, Moy L, Almeida EPRP, Woolen SA, Gichoya JW, Heye T, Hanneman K. Environmental Sustainability and AI in Radiology: A Double-Edged Sword. Radiology. 2024 Feb;310(2):e232030. doi: 10.1148/radiol.232030. PMID: 38411520; PMCID: PMC10902597.

2️⃣ Rengier F, Notohamiprodjo M, Weber MA. Thoughts on sustainability in the use of iodinated contrast media in CT: a practice-oriented review based on the example of a hospital and a private practice. Rofo. 2024 Feb 26;a-2246-6697.

3️⃣ Evan Halper, Amid explosive demand, America is running out of power, Washington Post, accessed 03-31-24 https://www.washingtonpost.com/business/2024/03/07/ai-data-centers-power/?utm_campaign=mb&utm_medium=newsletter&utm_source=morning_brew