In the age of digital transformation, federated learning (FL) has established itself as a key architecture for training AI models without compromising data privacy. However, their implementation in healthcare settings—such as the conversion of magnetic resonance imaging (MRI) to computed tomography (CT) scans—often requires such a high computational load that it excludes facilities with limited resources. This imbalance exacerbates existing gaps in health care. To address this, an innovative approach emerges: ecological federated learning with adaptive layer freezing, a method that reduces energy consumption and CO₂ emissions without sacrificing clinical accuracy.
The proposal is based on a mechanism of selective freezing of weights in the encoder of the neural network during federated training. Instead of updating all the parameters in each round, the relative difference in weights between rounds is monitored. Only when updates are consistently minimal—with adjustable patience—do the corresponding layers freeze. This eliminates the need to communicate and recalculate millions of redundant parameters, dynamically optimizing resource consumption. The results are compelling: reductions of up to 23% in training time, total energy and CO₂ equivalent emissions, measured with tools such as CodeCarbon. And most importantly, the performance in the MRI to CT conversion remains stable, with minimal variations in mean absolute error (MAE). In fact, in three of the five architectures evaluated, no statistically significant differences were observed, and in the other two significant improvements were recorded.
This development not only represents a technical breakthrough, but aligns with a new generation of climate-responsible AI solutions. In a context where the demand for computing for deep learning models doubles every few months, energy efficiency becomes an ethical and economic imperative. The adaptive freeze strategy allows institutions with modest infrastructure to participate in federated networks, democratizing access to cutting-edge diagnostic tools. In addition, by reducing the carbon footprint, the objectives of environmental sustainability and health equity are simultaneously addressed.
For companies looking to implement these types of systems, the key is to have a technology partner that understands both the complexity of federated learning and the scalability and security requirements. This is where our expertise in enterprise AI makes a difference. At Q2BSTUDIO we offer custom application development that integrates federated learning techniques, optimizing the use of resources without neglecting regulatory compliance. Our teams design modular architectures that adapt to any cloud infrastructure, whether in AWS and Azure cloud services, and manage cybersecurity at each layer of the system. In addition, we combine these capabilities with business intelligence services, such as Power BI, to monitor real-time performance metrics and emissions, enabling clinical teams to make informed decisions.
Adaptive freezing is just one example of how Green AI's principles can land on real projects. The methodology can be extrapolated to other areas of medical imaging, such as tumour segmentation or the detection of anomalies in X-rays, and even to applications outside the healthcare field, such as the analysis of financial data or the optimisation of industrial processes. In each case, the strategy involves a delicate balance between accuracy, privacy and sustainability, a balance that is achieved through intelligent monitoring and automation of training decisions.
For organizations that want to lead in this field, the path is not only to adopt new tools, but to develop a culture of efficiency. Customized software solutions allow freeze mechanisms to be customized according to the specific characteristics of each data set and each network. For example, instead of applying a generic policy, AI agents can be trained to learn to decide which layers to freeze in real time, maximizing energy savings without human intervention. These AI agents are integrated into the federated pipeline and adjust their criteria according to the evolution of the training, an approach that we are already exploring in our R+D laboratories.
In addition, the stability and performance obtained in MRI-to-CT tests demonstrate that it is possible to reduce reliance on specialized hardware. This opens the door for hospitals in rural areas or developing countries to collaborate in global federated networks, contributing anonymized data without the need to purchase supercomputers. The synergy between the public cloud – with its AWS and Azure cloud services – and artificial intelligence allows these architectures to be deployed with adjusted operating costs, also guaranteeing cybersecurity in the transmission of data between nodes.
However, the real added value lies in the ability to measure and communicate impact. Business intelligence tools, such as Power BI, help visualize energy savings per federated round, emissions evolution, and clinical performance in executive dashboards. This transparency is essential to justify investments and align projects with the ESG (environmental, social and governance) criteria that are increasingly demanded by investors and regulators.
In short, ecological federated learning with adaptive freezing represents a firm step towards a fairer, more sustainable and accessible artificial intelligence. At Q2BSTUDIO, we work to make these innovations a business reality, offering tailor-made applications that integrate everything from strategic consulting to production deployment. We are committed to helping organizations transform data into accurate diagnostics, without leaving an unnecessary footprint on the planet.
