Unifying and Personalizing Weakly-supervised Federated Medical Image Segmentation via Adaptive Representation and Aggregation (2304.05635v1)

Abstract: Federated learning (FL) enables multiple sites to collaboratively train powerful deep models without compromising data privacy and security. The statistical heterogeneity (e.g., non-IID data and domain shifts) is a primary obstacle in FL, impairing the generalization performance of the global model. Weakly supervised segmentation, which uses sparsely-grained (i.e., point-, bounding box-, scribble-, block-wise) supervision, is increasingly being paid attention to due to its great potential of reducing annotation costs. However, there may exist label heterogeneity, i.e., different annotation forms across sites. In this paper, we propose a novel personalized FL framework for medical image segmentation, named FedICRA, which uniformly leverages heterogeneous weak supervision via adaptIve Contrastive Representation and Aggregation. Concretely, to facilitate personalized modeling and to avoid confusion, a channel selection based site contrastive representation module is employed to adaptively cluster intra-site embeddings and separate inter-site ones. To effectively integrate the common knowledge from the global model with the unique knowledge from each local model, an adaptive aggregation module is applied for updating and initializing local models at the element level. Additionally, a weakly supervised objective function that leverages a multiscale tree energy loss and a gated CRF loss is employed to generate more precise pseudo-labels and further boost the segmentation performance. Through extensive experiments on two distinct medical image segmentation tasks of different modalities, the proposed FedICRA demonstrates overwhelming performance over other state-of-the-art personalized FL methods. Its performance even approaches that of fully supervised training on centralized data. Our code and data are available at https://github.com/LLMir/FedICRA.

Analysis of Federated Learning Methods for Ophthalmic Image Segmentation

The anonymous paper titled "Appendix 2418" presents a comparative analysis of various Federated Learning (FL) methods in the context of ophthalmic image segmentation, focusing on specific modalities such as fundus and OCTA (Optical Coherence Tomography Angiography). This research employs multiple datasets, which include Drishti-G, RIM-ONE-r, REFUGE and other datasets annotated through different techniques (e.g., scribble, point, block, and bounding box). The modalities and diverse annotation styles provide a complex testing ground for assessing the efficacy of FL approaches, given the inherent variability in resolution and device manufacturer specifications across data sources.

Significant numerical results are provided through the evaluation of segmentation performance using the HD95 metric. The analysis includes methods such as FedAvg, FedProx, FT, FedBN, FedAP, FedRep, MetaFed, FedICRA, and controls utilizing both weak and full CT (Centralized Training) settings. Among these, FedICRA emerges as a strong performer, with average HD95 values recorded as 9.64 for the optic disc (OD) and 10.11 for the optic cup (OC) across the datasets from various sites. In comparison, other FL methods such as MetaFed and FedRep show notable declines in performance.

Highlighted within the paper are FedICRA's capabilities, which demonstrate competitive advantages over both traditional centralized training approaches and alternative FL methods. For instance, FedICRA achieves a balance between generalization across diverse patient data distributions while maintaining segmentation precision, evident in its lower HD95 values compared to other methods. Furthermore, the visualization results underscore how FedICRA can effectively leverage sparse annotations, performing on par with models having access to full ground truth masks when applied to decentralized datasets.

The implications of this research span both practical and theoretical dimensions. Practically, the paper underscores the potential of FL to maintain data privacy while facilitating high-quality segmentation across distributed datasets, particularly valuable in healthcare domains where patient data confidentiality is paramount. Theoretically, the variation in performance between FL and centralized approaches encourages further exploration of federated optimization strategies, adaptive personalization techniques, and robust aggregation mechanisms to handle the heterogeneity intrinsic to real-world medical imaging datasets.

Looking forward, extending this line of inquiry could involve refining FL algorithms to better handle the imbalance and noisy annotations characteristic of medical datasets. There's also considerable scope for investigating the interoperability of these methods across different imaging techniques beyond ophthalmology, which would broaden the applicability and reliability of FL solutions in broader medical contexts. Continuing advances in this domain hold promise for enhancing precision medicine, with federated techniques potentially becoming a cornerstone for collaborative, privacy-preserving machine learning in clinical practice.