JMIR Publications

ABSTRACT

Background:

Federated learning (FL) is a decentralized approach to machine learning which is attracting attention as a training strategy that overcomes medical data privacy regulations and generalization of deep learning algorithms. In this study, we performed ultrasound (US) image analysis using FL to predict the benignity or malignancy of thyroid nodules.

Objective:

The goal of this study is to collect clinical data from several medical institutions and to validate the performance of FL by comparing FL and conventional deep learning using the collected data.

Methods:

We pooled 8,457 thyroid US images (5,375 malignant, 3,082 benign) from six institutions, and conducted FL using five deep learning networks (VGG19, ResNet50, ResNext50, SE-ResNet50, and SE-ResNext50). We also performed conventional deep learning on the pooled data and compared the results with those of FL. We used 20% of the dataset for internal validation and 100 US images from another institute for external validation.

Results:

For internal validation, the area under the receiver operating characteristic (AUROC) curve of the FL was between 78.88 and 87.56, and the AUC of the conventional deep learning was between 82.61 and 91.57. For external validation, the AUROC of the FL was between 75.20 and 86.72, and the AUROC curve of the conventional deep learning was between 73.04 and 91.04.

Conclusions:

We demonstrated that US image analysis using FL is feasible and that the performance of FL using decentralized data is comparable to that of conventional deep learning using pooled data. FL is a potentially useful method for analyzing medical images using deep learning while protecting patients’ personal information.