Rate-My-LoRA: Efficient and Adaptive Federated Model Tuning for Cardiac MRI Segmentation

TL;DR

This work addresses privacy-preserving cardiac MRI segmentation in federated settings with non-IID data across institutions. It introduces Rate-My-LoRA, which couples Low-Rank Adaptation (LoRA) with adaptive aggregation: adapters are lightweight, updates are $W = W_0 + BA$ (with B of size d x r and A of size r x k) and the server updates $W_t = W_{t-1} + ( sum_c w_t(c) |D^c| B^c_t A^c_t ) / ( sum_c |D^c| )$. The method yields superior in-client and cross-client performance versus prior LoRA-FL approaches, while reducing communication: adapter payloads around 1.8–3.6 MB per client vs full weights ~28 MB and bandwidth reductions up to 94% per round with about 15.5× per-epoch savings. This enables scalable, privacy-preserving model personalization for resource-constrained medical centers and improves generalization across heterogeneous datasets.

Abstract

Cardiovascular disease (CVD) and cardiac dyssynchrony are major public health problems in the United States. Precise cardiac image segmentation is crucial for extracting quantitative measures that help categorize cardiac dyssynchrony. However, achieving high accuracy often depends on centralizing large datasets from different hospitals, which can be challenging due to privacy concerns. To solve this problem, Federated Learning (FL) is proposed to enable decentralized model training on such data without exchanging sensitive information. However, bandwidth limitations and data heterogeneity remain as significant challenges in conventional FL algorithms. In this paper, we propose a novel efficient and adaptive federate learning method for cardiac segmentation that improves model performance while reducing the bandwidth requirement. Our method leverages the low-rank adaptation (LoRA) to regularize model weight update and reduce communication overhead. We also propose a \mymethod{} aggregation technique to address data heterogeneity among clients. This technique adaptively penalizes the aggregated weights from different clients by comparing the validation accuracy in each client, allowing better generalization performance and fast local adaptation. In-client and cross-client evaluations on public cardiac MR datasets demonstrate the superiority of our method over other LoRA-based federate learning approaches.

Figures (6)

• Figure 1: A demonstration of our FL scenario. it shows a scenario where the bandwidth limits the transfer of the model with full weights. Additionally, the dataset exhibits size imbalance; one hospital has half the patients compared to others.
• Figure 2: Demonstration of the highly non-IID data on federated learning. Due to different imaging equipment, the style of the images across clients differs, as in the red box.
• Figure 3: Overview of our efficient and adaptive FL method: the server contains a pretrained model and each client has a local adapter. The red, green and blue label in the MR image represents LVC, LVM, and RVC, respectively. In each FL round, only the adapters in each client are updated.
• Figure 4: Our Rate-My-LoRA method that utilizes on-client data for evaluating the performance gains/losses of the aggregated model and adaptive adjust the aggregation weight of each adapter to achieve overall performance.
• Figure 5: 3D visualization results of in-client evaluation. (a-d), (e-h), (i-l) are the ground truth, local training, FedPETuning and the Rate-My-LoRA method results from the client 1, 2 and 3, respectively.