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

TL;DR

The paper tackles inequitable access to federated learning in medical imaging by introducing a Green AI-driven, patience-based encoder freezing strategy that reduces energy and computation in MRI-to-CT translation. It evaluates multiple encoder–decoder architectures across a four-client FL setup, demonstrating up to ~23% reductions in training time, energy, and emissions while largely preserving image quality on unseen data. Key findings show most architectures maintain MAE with no significant PSNR/SSIM losses, and some even improve MAE, supporting the approach's practical viability. This work lays groundwork for sustainable, privacy-preserving FL in clinical settings and suggests directions for automated hyperparameter tuning and broader applicability.

Abstract

Federated Learning (FL) holds the potential to advance equality in health by enabling diverse institutions to collaboratively train deep learning (DL) models, even with limited data. However, the significant resource requirements of FL often exclude centres with limited computational infrastructure, further widening existing healthcare disparities. To address this issue, we propose a Green AI-oriented adaptive layer-freezing strategy designed to reduce energy consumption and computational load while maintaining model performance. We tested our approach using different federated architectures for Magnetic Resonance Imaging (MRI)-to-Computed Tomography (CT) conversion. The proposed adaptive strategy optimises the federated training by selectively freezing the encoder weights based on the monitored relative difference of the encoder weights from round to round. A patience-based mechanism ensures that freezing only occurs when updates remain consistently minimal. The energy consumption and CO2eq emissions of the federation were tracked using the CodeCarbon library. Compared to equivalent non-frozen counterparts, our approach reduced training time, total energy consumption and CO2eq emissions by up to 23%. At the same time, the MRI-to-CT conversion performance was maintained, with only small variations in the Mean Absolute Error (MAE). Notably, for three out of the five evaluated architectures, no statistically significant differences were observed, while two architectures exhibited statistically significant improvements. Our work aligns with a research paradigm that promotes DL-based frameworks meeting clinical requirements while ensuring climatic, social, and economic sustainability. It lays the groundwork for novel FL evaluation frameworks, advancing privacy, equity and, more broadly, justice in AI-driven healthcare.

Figures (13)

• Figure 1: Representation of the proposed patient-based adaptive layer freezing. The server is responsible for calculating the percentage difference in the aggregated encoder ($\rho_{\%}$) round by round. The encoder freezing condition is triggered when $\rho _{\%}$ remains consistently below the threshold ($\tau$) established by the final user for a minimum of $\mathcal{N}$ (patience parameter) consecutive rounds.
• Figure 2: Training time per local epoch for each participating client (Centres A-D), comparing the pre-freeze and post-freeze phases across the evaluated models.
• Figure 3: Performance comparison of the models trained with and without the proposed patience-based adaptive encoder freezing. A $p-value < 0.001$ is indicated by $***$.
• Figure S1: Simple UNet UNetRonnenberger – Minimum MAE difference case.
• Figure S2: Simple UNet UNetRonnenberger – Maximum MAE difference case.