MuCALD_SplitFed

This is the official code repository of the paper "MuCALD-SplitFed: Causal-Latent Diffusion for Privacy-Preserving Multi-Task Split-Federated Medical Image Segmentation."

---

Overview

MuCALD-SplitFed is a multi-task Split Federated Learning (SplitFed) framework that integrates causal representation learning and latent diffusion to enable stable, privacy-preserving, and accurate medical image segmentation across heterogeneous institutions and tasks.

Unlike standard FL/ SplitFed, which usually assume a single shared task and become unstable in realistic multi-task settings, MuCALD-SplitFed:

• learns causal latent factors that capture task-relevant structure,
• injects and denoises diffusion noise at split points to obfuscate private information,
• uses domain-adversarial alignment to suppress domain-specific leakage and improve cross-task generalization.

Experiments on five heterogeneous medical segmentation datasets show that the baseline SplitFed tends to be unstable or fails to converge in multi-task settings, whereas MuCALD-SplitFed achieves consistent performance gains and stronger privacy resilience against reconstruction attacks.

---

Key Contributions

MuCALD-SplitFed has the following key contributions:

1. Causal modelling in SplitFed latent space
   First framework to integrate causal representation learning directly at the SplitFed split points for multi-task medical segmentation.

2. Causal Representation & Diffusion Module (CRDM)

   • Exogenous encoder + Neural Structural Causal Model (Neural-SCM)
   • Latent diffusion with denoising UNet (LDDM) to inject and remove noise while preserving causal structure.

3. Domain-Adversarial Causal Alignment (DACA)

   • Gradient Reversal Layer (GRL) and domain discriminator for cross-client/domain invariance and reduced domain-specific leakage.

4. Empirical validation on 5 clients/ datasets

   • Blastocysts, HAM10K, FHPsAOPMSB, MosMed, and Kvasir-SEG, demonstrating improved segmentation metrics and degraded reconstruction quality at split points compared to Baseline SplitFed and strong personalized / multi-task FL baselines.

---

Installation

Create and activate environment

conda create -n mucald_splitfed python=3.9 -y
conda activate mucald_splitfed

Install the requirements

pip install -r requirements.txt

Datasets

We used following 5 datasets.

1. Blastocyst dataset [1]
2. HAM10K dataset [2]
3. FHPsAOPMSB dataset [3]
4. Mosmed Dataset [4]
5. Kvasir-SEG dataset [5]

Model Components

Following is the system architecture.

The proposed architecture contains 3 major components.

1. Causal Representation and Diffusion Module (CRDM)
2. Domain-Adversarial Causal Alignment (DACA)

Ablation studies

We also performed below ablation studies to better isolate the contribution of each component.

• CRDM only (without DACA)
• DACA only (without CRDM)
• Causal graph discovery disabled (Neural-SCM replaced with a standard encoder)
• Diffusion-based obfuscation disabled (diffusion disbaled)
• Forward noising disabled

All the results we achieved are added to the "Results_Summary" folder in this repository.

References

[1] Lockhart, Lisette, Parvaneh Saeedi, Jason Au, and Jon Havelock. "Multi-label classification for automatic human blastocyst grading with severely imbalanced data." In 2019 IEEE 21st International Workshop on Multimedia Signal Processing (MMSP), pp. 1-6. IEEE, 2019.

[1] Tschandl, Philipp, Cliff Rosendahl, and Harald Kittler. "The HAM10000 dataset, a large collection of multi-source dermatoscopic images of common pigmented skin lesions." Scientific data 5, no. 1 (2018): 1-9.

[3] Yaosheng Lu, Mengqiang Zhou, Dengjiang Zhi, Minghong Zhou, Xiaosong Jiang, Ruiyu Qiu, Zhanhong Ou, Huijin Wang, Di Qiu, Mei Zhong, et al., “The jnu-ifm dataset for segmenting pubic symphysis-fetal head,” Data Br, vol. 41, pp. 107904, 2022.

[4] Sergey P Morozov et al., “Mosmed data: data set of 1110 chest ct scans performed during the covid-19 epidemic,” Digit. Di-agn., vol. 1, no. 1, pp. 49–59, 2020.

[5] ebesh Jha et al., “Kvasir-seg: A segmented polyp dataset,” in Proc. MMM. Springer, 2020, pp. 451–462.

---

Citation

If you find this project useful in your research, please cite our paper:

@article{Mucald_SplitFed_Shiranthika_2026,
  title={MuCALD-SplitFed: Causal-Latent Diffusion for Privacy-Preserving Multi-Task Split-Federated Medical Image Segmentation},
  author={Chamani Shiranthika, Hadi Hadizadeh, Parvaneh Saeedi},
  Conference={IEEE International conference on Image Processing (ICIP 2026)},
  year={2026}
}