Federated approach for glioblastoma radiogenomic methylation classification

The privacy and data management advantages introduced by the federated learning (FL) paradigm make it the perfect solution for medical data science. Yet, there is a lack of studies exploring the use of this technology in the context of radiogenomic classification, with only one relevant study analyzing the technology in this field. Previous studies explored the use of centralized learning (CL) in radiogenomics with no actual strategy for privacy preservation during the deployment of the analyzed models. The work reported in this paper aims at showing the feasibility of a reusable centralized-to-federated conversion framework, bringing the federated adaptation of a centralized submission to the RSNA-MICCAI Brain Tumor Radiogenomic Classification as a first use case, showing that results are comparable between the two paradigms under same constraints. To achieve these goals this work explores Flower as a FL framework for ease of use and potential for customization, and adapts an existent RSNA-MICCAI challenge submission using a DenseNet121 classification model and the middle coronal slice of the FLAIR acquisitions from each patient as training set. Both the CL and FL pipelines are validated locally on accuracy, F1-score, Cohen's kappa and Area-Under-Curve (AUC), generated over repeated runs to have enough data to be properly tested with a Wilcoxon Rank Test. Across ten repeated runs each for an increasing number of clients, the federated DenseNet121 achieved validation metrics (accuracy, F1-score, Cohen’s kappa, and AUC) statistically indistinguishable from the centralized baseline for client counts of 3, 4, and 5 (Wilcoxon rank-sum p > 0.05). Only in the 2-client configuration did the FL setting exhibit significantly tighter metric distributions, thus significant differences in F1-scores and AUCs (p <= 0.05). This artifact was attributed to deterministic, class-balanced splitting and fixed validation ordering in that extreme partitioning scheme. These results, over all, show the comparability of the validation metrics between the centralized and federated implementation of the DenseNet121 model, laying the groundwork for a more extensive "centralized-to-federated" adapter framework.