💻 computer science

Med-DualLoRA: Local Adaptation of Foundation Models for 3D Cardiac MRI

The paper proposes Med-DualLoRA, a privacy-preserving federated framework that enhances 3D cardiac MRI disease detection by disentangling globally shared and local low-rank adaptations to achieve superior performance and communication efficiency across heterogeneous multi-center data.

Original authors: Joan Perramon-Llussà, Amelia Jiménez-Sánchez, Grzegorz Skorupko, Fotis Avgoustidis, Carlos Martín-Isla, Karim Lekadir, Polyxeni Gkontra

Published 2026-03-12

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Joan Perramon-Llussà, Amelia Jiménez-Sánchez, Grzegorz Skorupko, Fotis Avgoustidis, Carlos Martín-Isla, Karim Lekadir, Polyxeni Gkontra

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 you are trying to teach a group of doctors how to spot heart disease using MRI scans. You have a super-smart "Foundation Model" (let's call him Dr. AI) who has already read millions of medical books and knows the general rules of anatomy.

However, there's a catch:

Privacy: The hospitals (clients) can't send their patient data to a central server because of strict privacy laws.
Variety: Each hospital uses different MRI machines (Siemens, GE, Philips, etc.). A scan from a Siemens machine looks slightly different than one from a GE machine. This is like trying to recognize a face in a black-and-white photo versus a color photo; the details are there, but the "texture" is different.
Data Scarcity: Some hospitals have very few patients with heart disease, while others have many.

If you just ask Dr. AI to learn from all of them at once (Centralized Training), you get the best results, but it's illegal because of privacy. If you let each hospital train their own version of Dr. AI using only their local data, the small hospitals end up with "bad" doctors who are biased because they haven't seen enough examples.

This paper introduces a clever new method called Med-DualLoRA to solve this. Here is how it works, using simple analogies:

The Problem: The "One-Size-Fits-All" vs. "Total Isolation" Dilemma

The Old Way (Centralized): Imagine all the hospitals mailing their patient files to one giant library. The library builds the perfect doctor. Problem: Illegal due to privacy.
The Standard Federated Way: Imagine the library sends a blank notebook to every hospital. Each hospital writes down what they learned and sends the whole notebook back. The library averages them all. Problem: The notebooks are huge (too much data to send), and if Hospital A has 100 patients and Hospital B has 1, the average is dominated by Hospital A. The unique quirks of Hospital B get lost.
The "Low-Rank" Way (LoRA): Instead of rewriting the whole notebook, the hospitals only write a few sticky notes (small updates) and send those back. Problem: Even with sticky notes, if the hospitals have very different data, the "global" average sticky note still confuses the model. It tries to be everything to everyone and ends up being good at nothing specific.

The Solution: Med-DualLoRA (The "Two-Notebook" System)

The authors propose a system where Dr. AI carries two types of sticky notes (adapters) at the same time:

The "Global" Notebook (Shared Knowledge):
- This contains the universal rules of heart disease that apply to everyone, regardless of which machine took the scan.
- How it works: Every hospital learns these rules locally, then sends only this Global Notebook to the central server. The server mixes them all together to make a "Super Global Rulebook" and sends it back.
- Analogy: This is like the universal grammar of a language. Everyone agrees on the basic sentence structure.
The "Local" Notebook (Private Nuances):
- This contains the specific quirks of that particular hospital. Maybe their Siemens machine makes the heart look slightly brighter, or their patients are slightly older on average.
- How it works: The hospital learns this locally, but never sends it to the server. It stays private on their computer.
- Analogy: This is like a local accent. You understand the universal grammar, but you speak with a specific regional flavor.

How They Work Together

When a patient comes in for a scan at Hospital A:

The system uses the Global Notebook to understand the basic heart disease patterns.
It also applies the Local Notebook to adjust for the fact that Hospital A uses a Siemens machine.
The result is a diagnosis that is both smart (based on global knowledge) and accurate for that specific machine (based on local knowledge).

Why is this a big deal? (The Results)

Privacy: Since the "Local Notebook" never leaves the hospital, patient data stays safe.
Efficiency: They only send the "Global Notebook" back and forth. This is tiny compared to sending the whole AI model. It's like sending a 1-page summary instead of a 500-page book.
Performance:
- Standard methods failed when the data was messy (different machines, different patient counts). They became biased or confused.
- Med-DualLoRA kept the "Global" knowledge strong while respecting the "Local" differences.
- The Magic Trick: They found they only needed to update two specific layers of the AI (like changing just two chapters of the textbook) to get almost the same performance as updating the whole book. This saves even more time and data.

The Bottom Line

Think of Med-DualLoRA as a way to train a team of doctors who are all experts in the same field but work in different neighborhoods with different tools. Instead of forcing them to all think exactly the same way (which fails), or letting them work in total isolation (which is inefficient), they share a common core of knowledge while keeping their local specialties private.

This allows them to detect heart disease accurately across different hospitals and machines, without ever violating patient privacy or clogging up the internet with massive data transfers.

1. Problem Statement

The paper addresses the challenges of adapting Foundation Models (FMs) for 3D Cardiac Magnetic Resonance (CMR) disease detection in multi-center clinical settings. Key issues include:

Data Privacy & Centralization: Centralized fine-tuning on pooled clinical data is often infeasible due to privacy constraints.
Federated Learning (FL) Limitations: Standard FL for large models incurs prohibitive communication overhead due to the transmission of high-dimensional parameter updates.
Heterogeneity (Non-IID): Federated adaptation often fails under heterogeneous data distributions (different vendors, scanners, and patient demographics), leading to model bias toward dominant clients and suboptimal performance on minority sites.
PEFT Limitations: While Parameter-Efficient Fine-Tuning (PEFT) methods like LoRA reduce communication costs, standard federated LoRA approaches often struggle to balance global knowledge transfer with local client-specific adaptation, resulting in performance degradation compared to centralized training.

2. Methodology: Med-DualLoRA

The authors propose Med-DualLoRA, a client-aware, communication-efficient federated fine-tuning framework. The core innovation is the explicit disentanglement of globally shared knowledge and local client-specific adaptations using an additive decomposition of Low-Rank Adaptation (LoRA) modules.

Architecture & Mechanism

Dual LoRA Modules: For a linear projection $W$ $W$ in a frozen transformer backbone, the model injects two distinct LoRA components:
1. Global LoRA ( $A_g, B_g$ ): Represents transferable, site-agnostic features.
2. Local LoRA ( $A_{li}, B_{li}$ ): Captures client-specific characteristics (e.g., specific scanner artifacts or local disease prevalence).
Forward Pass: The output for client $i$ is computed as:
$f_i(x) = Wx + \frac{\alpha}{\sqrt{r}}(B_g A_g x) + \frac{\alpha}{\sqrt{r}}(B_{li} A_{li} x)$
Where $r$ is the rank and $\alpha$ is a scaling factor.
Federated Optimization Protocol:
- Local Training: Clients train both global and local parameters on their private data.
- Aggregation: Only the Global LoRA parameters are transmitted to the central server and aggregated (using weighted averaging, FedAvg).
- Privacy: Local LoRA parameters remain strictly on the client side and are never shared.
- Inference: Each client combines the aggregated global LoRA with their private local LoRA during inference.

Experimental Setup

Base Model: CineMA, a state-of-the-art 3D CMR Foundation Model.
Datasets:
- ACDC: 150 patients (controlled conditions).
- M&Ms (Combined): 543 patients from 5 institutions and 4 vendors (Siemens, GE, Philips, Canon), representing realistic clinical variability.
Task: Binary disease detection (Diseased vs. Non-diseased) using SAX cine frames.
Baselines: Centralized Full Fine-tuning, Centralized Head-only, Centralized LoRA, Federated Head-only, and Federated Standard LoRA.

3. Key Contributions

First Structural Disentanglement in 3D Medical FL: Introduces the first framework to explicitly separate global and local low-rank adaptations for 3D medical imaging FMs in a federated setting.
Med-DualLoRA Framework: Proposes a novel additive formulation that allows collaborative learning of global representations while preserving client-specific variations without sharing local data or parameters.
Comprehensive Analysis: Provides an in-depth analysis of fine-tuning strategies for 3D CMR FMs across multi-center datasets, highlighting the trade-offs between communication cost and performance.
Efficiency Validation: Demonstrates that adapting only a small subset of transformer blocks (specifically 2 blocks) is sufficient to achieve near-optimal performance, drastically reducing communication overhead.

4. Results

The method was evaluated on balanced accuracy, sensitivity, specificity, and F1-score across multiple vendors.

Performance Superiority:
- Med-DualLoRA achieved a Balanced Accuracy of 0.768 in the federated setting.
- This represents a statistically significant improvement over Federated Standard LoRA (0.693, $p=0.0371$ ) and Federated Head-only (0.514, $p<0.001$ ).
- Crucially, it restored Specificity (0.612), which was severely degraded in other federated baselines (0.417 for LoRA, 0.03 for Head-only), indicating better handling of class imbalance.
- Performance was comparable to Centralized Full Fine-tuning (0.745), effectively closing the gap between federated and centralized approaches.
Communication Efficiency:
- Pareto Frontier Analysis: Med-DualLoRA with only 2 transformer blocks (approx. 28 MB/round) achieved near-optimal performance (0.7465), placing it on the efficiency frontier.
- Full adaptation (12 blocks) required ~154 MB/round but offered diminishing returns compared to the 2-block configuration.
- This proves that the method significantly reduces communication costs while maintaining robustness.
Robustness: The method maintained competitive performance across all vendors, including smaller datasets (e.g., Canon), where standard federated approaches failed due to data scarcity and distribution shifts.

5. Significance

Scalable Clinical Deployment: Med-DualLoRA offers a practical solution for deploying medical FMs in privacy-sensitive, resource-constrained environments where data cannot be centralized.
Addressing Non-IID Data: By decoupling global and local adaptations, the method mitigates the bias toward dominant clients often seen in standard Federated Learning, ensuring fairer performance across diverse medical institutions.
Resource Optimization: The finding that adapting only a few transformer blocks yields high performance suggests that future medical FL systems can be highly efficient, reducing bandwidth requirements and training time without sacrificing diagnostic accuracy.
Future Direction: The work sets a precedent for structurally disentangling representations in medical FL, paving the way for more advanced aggregation strategies and multimodal adaptations.