FedDG: Federated Domain Generalization on Medical Image Segmentation via Episodic Learning in Continuous Frequency Space

Abstract

Federated learning allows distributed medical institutions to collaboratively learn a shared prediction model with privacy protection. While at clinical deployment, the models trained in federated learning can still suffer from performance drop when applied to completely unseen hospitals outside the federation. In this paper, we point out and solve a novel problem setting of federated domain generalization (FedDG), which aims to learn a federated model from multiple distributed source domains such that it can directly generalize to unseen target domains. We present a novel approach, named as Episodic Learning in Continuous Frequency Space (ELCFS), for this problem by enabling each client to exploit multi-source data distributions under the challenging constraint of data decentralization. Our approach transmits the distribution information across clients in a privacy-protecting way through an effective continuous frequency space interpolation mechanism. With the transferred multi-source distributions, we further carefully design a boundary-oriented episodic learning paradigm to expose the local learning to domain distribution shifts and particularly meet the challenges of model generalization in medical image segmentation scenario. The effectiveness of our method is demonstrated with superior performance over state-of-the-arts and in-depth ablation experiments on two medical image segmentation tasks. The code is available at https://github.com/liuquande/FedDG-ELCFS.

Domain Generalization on Medical Imaging Classification using Episodic Training with Task Augmentation

“…The others design heuristic schemes, such as manipulating deep neural network architectures [37], leveraging self-supervision signals [13], designing novel training paradigm [38], [39], or conducting data augmentation [14], [40]. Very recently, with the rise of model-agnostic metalearning (MAML) [19], an paradigm of episodic learning is introduced to address DG [16], [17], [41]- [43]. The principle behind this is to simulate virtual meta-task episodically so that the knowledge of maintaining generalization under domain shift can be transferred to the real task of DG.…”

“…Simple data augmentation as random flipping is applied. Following the common practice in DG [13], [16], [17], we use leave-one-domain-out cross-validation, i.e., training on multiple source domains and testing on the remaining unseen one, with randomly dividing the samples of each domain into 70% training data and 30% testing data. To evaluate the performance on epithelium-stroma classification, we adopt the accuracy metric as ACC = Nc Nt , where N c and N t represents the number of correctly classified images and total number of images, respectively.…”

“…In contrast, it has a broad prospect for clinical practice to train a generalizable model that can be directly applied to a new domain (cf. Figure 1), i.e., domain generalization (DG) [12]- [17]. Without accessing the information of test domain, the principle of DG is to capture the representation that is insensitive to the changes of domain-specific statistics.…”

“…Recently, model-agnostic meta-learning (MAML) [19], which aims to train the model that can be fast adapted to new tasks, has been promising in many applications like few-shot learning. Then this spirit is introduced to address DG with a paradigm of episodic training [15]- [17]. A key difference between episodic training and conventional supervised training is that the data point in episodic training is a task.…”

Subgraph Federated Learning with Missing Neighbor Generation

“…Federated learning (FL) [5,6], targeting at training machine learning models with data distributed in multiple local systems to resolve the information-silo problem, has shown its advantage in enhancing the performance and generalizability of the collaboratively trained models without the need of sharing any actual data. For example, FL has been devised in computer vision (CV) and natural language processing (NLP) to allow the joint training of powerful and generalizable deep convolutional neural networks and language models on separately stored datasets of images and texts [7,8,9,10,11].…”