Federated Disentangled Representation Learning for Unsupervised Brain Anomaly Detection

Abstract

Recent advances in Deep Learning (DL) and the increased use of brain MRI have provided a great opportunity and interest in automated anomaly segmentation to support human interpretation and improve clinical workflow. However, medical imaging must be curated by trained clinicians, which is time-consuming and expensive. Further, data is often scattered across multiple institutions, with privacy regulations limiting its access. Here, we present FedDis (Federated Disentangled representation learning for unsupervised brain pathology segmentation) to collaboratively train an unsupervised deep convolutional neural network on 1532 healthy MR scans from four different institutions, and evaluate its performance in identifying abnormal brain MRIs including multiple sclerosis (MS) lesions, low-grade tumors (LGG), and high-grade tumors/glioblastoma (HGG/GB) on a total of ~500 scans from 5 different institutions and datasets. FedDis mitigates the statistical heterogeneity given by different scanners by disentangling the parameter space into global, i.e., shape and local, i.e., appearance. We only share the former with the federated clients to leverage common anatomical structure while keeping client-specific contrast information private. We have shown that our collaborative approach, FedDis, improves anomaly segmentation results by 99.74% for MS and 40.45% for tumors over locally trained models without the need for annotations or sharing private local data. We found out that FedDis is especially beneficial for clients that share both healthy and anomaly data coming from the same institute, improving their local anomaly detection performance by up to 227% for MS lesions and 77% for brain tumors.

A tool for federated training of segmentation models on whole slide images

“…Numerous examples of federated learning on medical data exist, 26 , 27 , 28 , 29 however at this time computational pathology research on federated learning using WSIs is limited to a paper by Lu et al 30 Lu et al trained a weakly supervised, multi instance learning model for subtyping breast cancer and renal cell carcinoma and predicting survival, while exploring the effects of differential privacy 31 on model performance. Setting aside the complexities of network hyperparameter tuning, we argue that federated learning is a data organization and synchronization problem at its core.…”

Evaluation of pseudo-healthy image reconstruction for anomaly detection with deep generative models: Application to brain FDG PET

“…More recently, following the success of diffusion models for image generation, DDPMs have also been used for anomaly detection tasks in medical imaging (Wolleb et al, 2022;Pinaya et al, 2022a;Bercea et al, 2023a).…”

“…In most of the other cases, when the ground truth anomaly mask is not available, the evaluation consists in applying a classifier to the reconstructed images that was trained to distinguish pathological and healthy images, or using the reconstruction error itself from which an anomaly score is derived. One way to improve the evaluation is to use synthetic data by corrupting real healthy data with sprites (Bercea et al, 2023b;Pinaya et al, 2022b).…”

Cross-Modal Federated Human Activity Recognition via Modality-Agnostic and Modality-Specific Representation Learning

“…FedBN (Li et al 2021b) uses local batch normalization to alleviate the feature shift between different clients. FedDis (Bercea et al 2021) is a FL method for unsupervised brain pathology segmentation, which can disentangle the parameter space into shape and appearance. Fed-CMR (Zong et al 2021) focuses on a federated cross modal retrieval task where each client has both text and image data.…”