ACDN Federated Learning Server

This site provides project and connectivity information for the Australian Cancer Data Network (ACDN) federated learning studies.
Purpose Provider Architecture FL Communication Networking Contacts

Purpose

This project is part of ACDN and enables multiple hospitals to collaboratively develop machine learning models while maintaining patient privacy. It uses Federated Learning (FL): each hospital trains locally on its de-identified datasets, while only numeric model updates are exchanged with a central server for aggregation.

No Patient Health Information (PHI) is transferred to the FL server. Raw clinical datasets remain within each hospital.

Who provides flserver.australian-cancer-data.network?

Where are the server and data located?

Data sources and architecture

Each participating hospital (under its Local Health District, LHD) identifies cohorts (e.g., breast or lung cancer) and performs local extraction from clinical systems.

De-identification and separation of identifiable data

Identifiable patient information is managed within a dedicated local identification-key environment (separate VM/system). Access is restricted to authorised personnel according to local governance.

Identifiable data required for linkage is processed through an anonymisation pipeline using the RSNA MIRC Clinical Trials Processor (CTP) to produce de-identified datasets for research use.

ACDN hospital research environment

Following de-identification, research datasets are stored within a separate research VM/environment at each site:

Researchers typically access services via controlled interfaces and secure ports with authentication enforced at multiple layers.

Figure 1: De-Identified Data Extraction in ACDN Systems under LHDs
Figure 1: De-Identified Data Extraction in ACDN Systems under LHDs

Federated Learning (FL)-based model communication

Once de-identified datasets are prepared and stored locally within each hospital, Federated Learning is used to collaboratively train models without transferring patient data outside the hospital network.

What is sent from each client to the FL server

What is NOT sent

No Patient Health Information (PHI) is transferred. Raw clinical datasets remain within each hospital. The FL server performs aggregation of model updates only.

Federated learning server and client interaction

Figure 2: Federated Learning communication between global server and hospital clients
Figure 2: Federated Learning communication between a global server and hospital clients

Networking requirements (domain + port 5010)

FL client communication from participating hospital research VMs may be restricted by local proxy/firewall policies. For federated learning participation, sites may need to allow outbound connectivity to a destination domain and port.

The requested whitelisting enables encrypted model-update exchange between clients and the server without transmitting PHI. Additional servers/domains may be used for externally run projects as needed, following the same principle: only model updates are exchanged.

Contacts