🔬 physics

Energy-Efficient Federated Learning via Adaptive Encoder Freezing for MRI-to-CT Conversion: A Green AI-Guided Research

This paper proposes a Green AI-guided adaptive encoder freezing strategy for federated learning in MRI-to-CT conversion that significantly reduces energy consumption and CO2 emissions by up to 23% while maintaining or improving model performance, thereby promoting equitable and sustainable healthcare AI.

Original authors: Ciro Benito Raggio, Lucia Migliorelli, Nils Skupien, Mathias Krohmer Zabaleta, Oliver Blanck, Francesco Cicone, Giuseppe Lucio Cascini, Paolo Zaffino, Maria Francesca Spadea

Published 2026-07-17

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Ciro Benito Raggio, Lucia Migliorelli, Nils Skupien, Mathias Krohmer Zabaleta, Oliver Blanck, Francesco Cicone, Giuseppe Lucio Cascini, Paolo Zaffino, Maria Francesca Spadea

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine a world where hospitals could share their medical secrets to build a super-smart AI doctor, but without ever actually sharing the patients' private files. This is the promise of Federated Learning. Think of it like a group of chefs from different countries trying to perfect a secret soup recipe. Instead of sending their ingredients across the ocean (which is slow and risky), they each cook a little bit of the soup in their own kitchens, send only the taste notes to a central judge, and the judge mixes those notes to create a master recipe. Everyone gets better at cooking, but no one ever sees the other's secret ingredients.

However, there's a catch. Cooking these complex soups takes a lot of energy, heat, and expensive equipment. In the tech world, this is often called "Red AI"—it's powerful, but it burns a lot of fuel and creates a huge carbon footprint. This creates a problem: only the richest hospitals with the fanciest kitchens can afford to join the club, leaving smaller clinics behind. Enter Green AI, a movement that asks, "Can we make these smart systems work just as well, but with less energy and less waste?" The goal is to make medical AI fair, affordable, and friendly to the planet, ensuring that a small clinic in a rural town can learn from a big city hospital without needing a supercomputer to do it.

The Paper's Big Idea: Freezing the Brain to Save the Planet

This paper tackles a specific challenge in medical imaging: turning MRI scans (which look like soft, gray blobs) into CT scans (which show sharp, bone-white details). Doctors need these "fake" CT scans to plan radiation therapy, but making them usually requires a massive, energy-hungry AI model. The researchers asked: Can we make this process faster and greener without losing accuracy?

Their solution is a clever trick they call Adaptive Encoder Freezing. To understand this, imagine the AI model as a two-part machine: a Scanner (the encoder) that looks at the MRI and figures out "what is this?" and a Painter (the decoder) that draws the final CT image.

In a standard Federated Learning setup, every time the chefs (hospitals) send their taste notes, the whole machine has to be recalibrated. It's like rewiring the entire kitchen every time someone adds a pinch of salt. The researchers realized that once the "Scanner" part of the machine has figured out the basics of what the MRI looks like, it doesn't need to keep changing its mind. It just needs to stay steady while the "Painter" part does the fine-tuning.

So, they invented a patience-based mechanism. Here's how it works:

The central server watches the "Scanner" weights (the settings that decide how the AI sees the image) as the training goes on.
It calculates how much the settings change from one round to the next.
If the settings stop changing much (meaning the AI has "learned" the basics), the server says, "Okay, we're being patient. Let's wait a few more rounds to be sure."
If the settings stay calm for a specific number of rounds (they used 3 rounds as a test), the server hits the "Freeze" button on the Scanner.

Once frozen, the Scanner stops doing the heavy lifting. The AI only updates the Painter. This is like telling the head chef to stop worrying about the recipe's base and just focus on the garnish. Because the Scanner is frozen, the computers at the hospitals don't have to do as much math, saving a massive amount of time and electricity.

What They Found

The team tested this idea on five different AI architectures (different types of model designs) using data from four different hospitals (Centres A, B, C, and D) and validated it on a fifth (Centre E). They ran the experiments five times to make sure the results were real and not just luck.

The results were quite promising:

Time and Energy: By freezing the encoder, they reduced the total training time by up to 22.0% (for one specific model). This directly translated to energy savings. The total energy consumption dropped by up to 23.0%, and the carbon emissions (measured in kgCO2eq) dropped by up to 23.2%.
Accuracy: The most important question was, "Did the frozen brain make worse pictures?" The answer was generally no. For three out of the five models they tested, the quality of the generated CT scans was statistically the same as the non-frozen versions. For the other two models, the frozen versions actually showed a tiny, statistically significant improvement in accuracy (measured by Mean Absolute Error), though the difference was so small (about 2-3 HU) that it wouldn't matter to a doctor planning treatment.
Consistency: The visual quality, measured by metrics like PSNR and SSIM, remained almost identical across the board. The "frozen" models didn't produce blurry or weird images; they just got there faster.

Why This Matters

The paper suggests that this "patience-based freezing" is a smart way to make Federated Learning more accessible. By cutting the computational load, smaller hospitals with less powerful computers can now participate in these collaborative AI projects without burning a hole in their budget or the planet's atmosphere.

The authors are careful to note that while their specific settings (a 5% threshold for change and 3 rounds of patience) worked well for these specific MRI-to-CT models, they aren't a magic wand for every single AI in existence. Other types of models might need different settings. However, the core idea—that we can pause the parts of an AI that have already learned their job to save energy—is a solid step toward "Green AI." It proves that we don't have to choose between powerful medical tools and a sustainable future; with a little bit of smart engineering, we can have both.

Technical Summary: Energy-Efficient Federated Learning via Adaptive Encoder Freezing for MRI-to-CT Conversion

Problem Statement
Federated Learning (FL) offers a pathway to collaborative deep learning (DL) in healthcare without centralizing sensitive patient data, potentially addressing disparities in access to AI-driven diagnostics. However, the computational and energy demands of training complex DL models often exclude resource-constrained medical centers, thereby reinforcing existing healthcare inequalities. This "Red AI" paradigm prioritizes performance at the cost of sustainability. While Green AI advocates for sustainable practices, existing FL optimization strategies—such as communication reduction via increased local epochs or model compression—often fail to sufficiently reduce the local computational burden or may degrade model accuracy. Furthermore, prior layer-freezing studies have largely focused on communication efficiency or shallow networks, lacking application to deep encoder-decoder architectures in medical image-to-image (I2I) translation tasks like MRI-to-Computed Tomography (CT) conversion.

Methodology
The study proposes a patience-based adaptive encoder freezing strategy designed to reduce computational load and energy consumption in FL settings while maintaining model performance. The approach is applied to five distinct encoder-decoder architectures (including U-Net variants) used for MRI-to-CT synthesis.

Adaptive Freezing Mechanism: The method selectively freezes the weights of the encoder portion of the network. Freezing is not immediate but is triggered dynamically based on the convergence of the global model.
Convergence Metric: The server monitors the relative percentage difference ( $\rho\%$ ) of the aggregated encoder weights between consecutive training rounds. This metric is calculated using the Mean Absolute Error (MAE) between corresponding encoder layers of the current and previous rounds.
Patience-Based Trigger: A freezing condition is activated only when $\rho\%$ remains below a user-defined threshold ( $\tau$ ) for a consecutive number of rounds ( $N$ , the patience parameter). This prevents premature freezing due to transient fluctuations.
Experimental Setup: The framework was tested in a simulated FL environment involving four clinical centers (clients) and one server. The experiments utilized brain MRI data from five distinct centers (A–E), with Center E serving as an unseen test set. The study tracked training time, energy consumption (kWh), and CO2 equivalent emissions using the CodeCarbon library. The best-performing aggregation strategy from previous work (FedAvg combined with FedProx) was used, with a fixed 25 communication rounds.

Key Results
The proposed adaptive freezing strategy was evaluated across five architectures. The results demonstrated significant efficiency gains without compromising the clinical utility of the synthetic CTs:

Efficiency Gains: The method reduced total training time, energy consumption, and CO2eq emissions by up to 23.2% (specifically for the Li et al. [36] architecture). Across all models, energy and emission reductions ranged from 9.1% to 23.2%, with an average reduction of approximately 19.9%.
Training Time: Global training time decreased by between 9.4% and 22.0% depending on the architecture. Per-epoch training times for local clients were also significantly reduced (e.g., up to 35.5% reduction for specific models on specific centers).
Model Performance:
- MAE: For three of the five architectures (Simple U-Net, Fu et al., and Li et al. 2020), there were no statistically significant differences in Mean Absolute Error (MAE) compared to non-frozen counterparts. For the Li et al. [36] and Spadea et al. [37] architectures, statistically significant improvements in MAE were observed ( $p < 0.001$ ), though the absolute reduction (approx. 2–3 Hounsfield Units) remained well below radiologically meaningful thresholds (20–50 HU).
- PSNR and SSIM: No systematic differences were observed in Peak Signal-to-Noise Ratio (PSNR) or Structural Similarity Index Measure (SSIM) between frozen and non-frozen models.
Generalization: The models maintained their ability to generalize to the unseen dataset (Center E), confirming that freezing the encoder did not hinder the model's capacity to adapt to local data heterogeneity via the decoder.

Significance and Claims
The paper positions this work as an operationalization of Green AI and equity in Federated Learning. By reducing the computational and energy barriers to entry, the proposed method aims to make FL more accessible to resource-limited institutions, thereby preventing the concentration of AI benefits in well-resourced centers.

The authors claim that their approach:

Balances Efficiency and Performance: It achieves substantial reductions in environmental impact and training time while preserving (and occasionally marginally improving) the fidelity of medical image synthesis.
Addresses a Gap: Unlike previous freezing strategies focused on communication or memory, this work explicitly targets energy efficiency and scalability in deep medical I2I translation tasks.
Promotes Sustainability: It aligns with the "Green-in AI" paradigm by designing systems that are sustainable by default, reducing the carbon footprint of clinical AI deployment.

The study acknowledges that the selection of hyperparameters ( $\tau$ and $N$ ) requires preliminary observation of weight dynamics for specific models, noting this as a limitation and a direction for future automated tuning. However, the proposed configuration ( $\tau=5\%, N=3$ ) is presented as a robust baseline for the evaluated architectures.

The Paper's Big Idea: Freezing the Brain to Save the Planet

What They Found

Why This Matters

More like this