Computer Science editorial
Open AccessOA2026
FedMust: Semi-supervised Multi-task Student-Teacher Federated Learning for Multi-organ CT Segmentation
A federated student-teacher framework that exploits unlabeled data and multi-task learning to boost multi-organ CT segmentation across privacy-constrained institutions.
Ashkan Moradi; Bendik Skarre Abrahamsen; Mattijs Elschotยท 2026ยท DOI 10.48550/arXiv.2609.24627
The core problem
Multi-organ segmentation from CT scans is a cornerstone of modern clinical workflows, enabling diagnosis, treatment planning, and quantitative analysis. Deep learning models for this task, however, demand large volumes of annotated patient data. In practice, individual institutions rarely possess sufficiently large and diverse annotated datasets, and privacy regulations prevent them from sharing patient data to overcome this limitation. Furthermore, annotation is labor-intensive and requires scarce multi-organ expertise, so institutions typically label only a small fraction of their local data, leaving the larger unlabeled portion unused. This work proposes FedMust, a flexible semi-supervised federated multi-task student-teacher framework that leverages federated learning (FL) to improve multi-organ segmentation by exploiting both labeled and unlabeled data across participating sites. The framework is designed to relax labeled-data requirements for client participation while still benefiting from cross-site collaboration.
Innovation
Extensive experiments demonstrated the effectiveness of FedMust compared with local and federated single-organ models, yielding an average performance gain of 13 percent across clients. The experiments also demonstrated the impact of multi-task learning and unlabeled data, and the applicability of the framework in relaxing labeled-data requirements for client participation. These results indicate that FedMust can effectively utilize unlabeled data and multi-task supervision to improve segmentation performance in federated settings where labeled data are scarce and privacy constraints prevent data sharing. The code is available at https://github.com/AshknMrd/FedMust.
Multi-organ segmentation from CT scans is a cornerstone of modern clinical workflows, enabling diagnosis, treatment planning, and quantitative analysis. Deep learning models for this task, however, demand large volumes of annotated patient data. In practice, individual institutions rarely possess sufficiently large and diverse annotated datasets, and privacy regulations prevent them from sharing patient data to overcome this limitation. Furthermore, annotation is labor-intensive and requires scarce multi-organ expertise, so institutions typically label only a small fraction of their local data, leaving the larger unlabeled portion unused. This work proposes FedMust, a flexible semi-supervised federated multi-task student-teacher framework that leverages federated learning (FL) to improve multi-organ segmentation by exploiting both labeled and unlabeled data across participating sites. The framework is designed to relax labeled-data requirements for client participation while still benefiting from cross-site collaboration.
FedMust operates in communication rounds. At each round, local training begins with clients that have labels for the same task forming a federation to produce an aggregated teacher model. Formally, for a task , the teacher parameters are obtained by federated averaging over the set of clients that possess labels for task :
Why it matters
The FedMust framework addresses two critical challenges in medical imaging: limited annotated data and privacy constraints. By combining semi-supervised learning with federated multi-task learning, it enables institutions to collaborate without sharing patient data, while leveraging unlabeled data that would otherwise be discarded. The student-teacher architecture allows knowledge distillation from task-specific teachers to a multi-task student, promoting feature sharing across tasks and improving generalization. The 13 percent average performance gain across clients underscores the benefits of multi-task learning and unlabeled data. Furthermore, the framework relaxes labeled-data requirements, as clients without labels for a specific task can still participate in the student federation and contribute unlabeled data. This flexibility could facilitate broader participation in federated learning initiatives, particularly from smaller institutions with limited annotation resources. Future work may explore scalability to more tasks and organs, robustness to heterogeneous data distributions, and integration with other semi-supervised techniques.
Who should read this
CS practitioners and researchers
Opening member contentโฆ