Edition 7 - AI’s Role in Sustainable Healthcare: Radiology, Nuclear Medicine, & Beyond

Paul’s Thoughts:

Deep learning, which is utilised by many healthcare devices, is a computationally heavy process with a high carbon footprint. There is a growing concern that it may accelerate climate change and there is a need to mitigate its effects. This review assessed the use of AI in heathcare in terms of climate change. The review assessed that, actually, AI may reduce climiate change as it will allow for a smoother, faster and less wasteful workflow. From the incerased use of telemedicine and thus less travel for patients, to improving identification and reducing the need for repeat examinations, the reduction of unnecessary patient travel will bring down greenhouse gas emissions. However, AI models are energy-intensive, with one study reporting that a single large AI model takes as much energy to run as five cars over their entire lifespan. The use of AI in healthcare depends on data centers that use servers, cooling systems, and networking platforms. All these must run constantly in controlled environments, consuming a lot of energy and accounting for about 1% of global power consumption. Additionally, hardware updates produce large amounts of electronic waste, as the materials of lead, cadmium and mercury are commonly used. Possible solutions suggested include increasing the energy efficiency of AI models via techniques such as quantisation and pruning. Improved infrastructure design, revamping hardware and software concepts, and efficient power management using dynamic voltage and frequency scaling can also reduce AI's environmental costs. The review also states that collaboration among stakeholders and government, if initiated at an early stage, is critical for these trends to become the norm.

Timescale: Mid | 5 Years

Paul’s Thoughts:

This is an important, but widely not discussed, topic of AI model development and implementation. The authors make a number of good suggestions to minimise the environmental impact of radiology AI applications. They suggest the use of low-power CPUs and multi-institution collaboration with centralised data sharing and federated learning models. Data compression to minimise data storage was recommended, as well as tiered storage systems based on the frequency of access. Partnerships with cloud-service providers that are committed to renewable energy sources and energy-conscious service scheduling. Additionally, training AI models in locations with renewable energy sources and in cooler climates to minimise the energy needed for cooling were suggested. This paper is an important milestone that should act as a guiding post for AI model developers and large institutions wishing to implement AI in radiology applications.

Timescale:  Early | 3 Years 

Paul’s Thoughts:

It was suggested that generative AI has the potential to enhance communication with patients, patient information and student learning. ChatGPT powered by GPT-4 and using DALL-E−3 was asked to generate an image for teaching purposes that demonstrated: a nuclear medicine bone scan with skeletal metastases; coronary artery disease; an F-18 FDG-PET scan of the brain with Alzheimer's disease; and an x-ray showing a fracture. These were concluded to be misleading for educational purposes and for patients who may be searching for information, suggesting that AI is not yet ready for such use.

Timescale: Early | 2 Years