The growing need to apply deep learning techniques to sensitive clinical data clashes with privacy and regulatory constraints that prevent data centralization. This thesis proposes and validates an end-to-end framework based on Federated Learning (FL) for Hospital Data Spaces. The architecture adopts a systemic perspective, separating coordination (Orchestrator) and execution (Data Space Connectors) through distributed microservices. The experimental validation, conducted on the BraTS 2020 dataset partitioned to simulate a federation of three hospitals with limited resources and non-IID data, demonstrates the system's ability to orchestrate distributed training cycles without centralizing data. The U-Net segmentation model achieved a Dice Score of 0.715, while the CNN for clinical marker extraction achieved an R² of 0.34, confirming the framework's resilience under realistic conditions. The main contributions are: the empirical validation of FL in simulated clinical scenarios with heterogeneous data, the design and implementation of an operational and reproducible framework, and the translation of technical solutions into responses to real-world constraints that hinder clinical adoption. These results provide a solid foundation for future developments toward multi-center deployments, strengthening FL as a viable and ethical solution for collaboration in the clinical imaging domain.
La crescente necessità di applicare tecniche di deep learning a dati clinici sensibili si scontra con vincoli di privacy e normative che impediscono la centralizzazione dei dati. Questa tesi propone e valida un framework end-to-end basato su Federated Learning (FL) per contesti di Hospital Data Spaces. L’architettura adotta una prospettiva sistemica, separando coordinamento (Orchestrator) ed esecuzione (Data Space Connectors) tramite microservizi distribuiti. La validazione sperimentale, condotta sul dataset BraTS 2020 partizionato per simulare una federazione di tre ospedali con risorse limitate e dati non-IID, dimostra la capacità del sistema di orchestrare cicli di training distribuiti senza centralizzare i dati. Il modello di segmentazione U-Net ha raggiunto un Dice Score di 0.715, mentre la CNN per l’estrazione di marker clinici ha ottenuto un R² di 0.34, confermando la resilienza del framework in condizioni realistiche. I principali contributi sono: la validazione empirica del FL in scenari clinici simulati con dati eterogenei, la progettazione e implementazione di un framework operativo e riproducibile, e la traduzione di soluzioni tecniche in risposte a vincoli reali che ostacolano l’adozione clinica. Questi risultati costituiscono una base solida per futuri sviluppi verso deployment multicentrici, rafforzando il FL come soluzione praticabile ed etica per la collaborazione nel dominio dell’imaging clinico.
FEDERATED LEARNING PER HOSPITAL DATA SPACE
MARSIGLIA, ALESSIO
2024/2025
Abstract
The growing need to apply deep learning techniques to sensitive clinical data clashes with privacy and regulatory constraints that prevent data centralization. This thesis proposes and validates an end-to-end framework based on Federated Learning (FL) for Hospital Data Spaces. The architecture adopts a systemic perspective, separating coordination (Orchestrator) and execution (Data Space Connectors) through distributed microservices. The experimental validation, conducted on the BraTS 2020 dataset partitioned to simulate a federation of three hospitals with limited resources and non-IID data, demonstrates the system's ability to orchestrate distributed training cycles without centralizing data. The U-Net segmentation model achieved a Dice Score of 0.715, while the CNN for clinical marker extraction achieved an R² of 0.34, confirming the framework's resilience under realistic conditions. The main contributions are: the empirical validation of FL in simulated clinical scenarios with heterogeneous data, the design and implementation of an operational and reproducible framework, and the translation of technical solutions into responses to real-world constraints that hinder clinical adoption. These results provide a solid foundation for future developments toward multi-center deployments, strengthening FL as a viable and ethical solution for collaboration in the clinical imaging domain.I documenti in UNITESI sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/20.500.14253/70655