Federated Learning for Thyroid Ultrasound Image Analysis to Protect Personal Information: Validation Study in a Real Health Care Environment

Abstract

Background Federated learning is a decentralized approach to machine learning; it is a training strategy that overcomes medical data privacy regulations and generalizes deep learning algorithms. Federated learning mitigates many systemic privacy risks by sharing only the model and parameters for training, without the need to export existing medical data sets. In this study, we performed ultrasound image analysis using federated learning to predict whether thyroid nodules were benign or malignant. Objective The goal of this study was to evaluate whether the performance of federated learning was comparable with that of conventional deep learning. Methods A total of 8457 (5375 malignant, 3082 benign) ultrasound images were collected from 6 institutions and used for federated learning and conventional deep learning. Five deep learning networks (VGG19, ResNet50, ResNext50, SE-ResNet50, and SE-ResNext50) were used. Using stratified random sampling, we selected 20% (1075 malignant, 616 benign) of the total images for internal validation. For external validation, we used 100 ultrasound images (50 malignant, 50 benign) from another institution. Results For internal validation, the area under the receiver operating characteristic (AUROC) curve for federated learning was between 78.88% and 87.56%, and the AUROC for conventional deep learning was between 82.61% and 91.57%. For external validation, the AUROC for federated learning was between 75.20% and 86.72%, and the AUROC curve for conventional deep learning was between 73.04% and 91.04%. Conclusions We demonstrated that the performance of federated learning using decentralized data was comparable to that of conventional deep learning using pooled data. Federated learning might be potentially useful for analyzing medical images while protecting patients’ personal information.

A Review of Medical Federated Learning: Applications in Oncology and Cancer Research

“…The resulting model was able to better identify pancreas from both datasets than models trained only on one site and validated on the other. Concluding with similar results, [53] tested several deep learning architectures for federated thyroid images classification, and Choudhury et al [54] used data from 3 different sites to create a prediction model for patients with anal cancer, an extremely rare form of cancer, who received radical chemoradiotherapy. The large and diverse group of examples given here demonstrates the robustness and versatility of the Federated Learning paradigm, as well as its ability to improve automated analysis on more rare cancer cases [51,53,54].…”

Distributed Anomaly Detection in Smart Grids: A Federated Learning-Based Approach

“…In the proposed approach, multiple hospitals collaboratively trained a CNN model without transferring their private data. Lee et al [23] used FL for analysis of thyroid ultrasound images. The performance of the FL approach was found to be comparable with centralised training with enhanced privacy and significant savings in bandwidth.…”

A Systematic Review of Federated Learning in the Healthcare Area: From the Perspective of Data Properties and Applications

“…HFL data partition is quite common in FL applied for medical applications. More than half of FL studies on medical applications implemented horizontal medical data partition in their experiment [18,19,21,37,[39][40][41][42][43][44][45][46][47][48][49]51,52,54,55]. Unlike FL applied for nonmedical applications where training is carried out across many nodes, FL studies in medical applications only handle limited nodes from 2 to 100, as listed in Table 2.…”

“…In machine learning, classification algorithms learn how to classify or annotate a given set of instances with classes or labels. There are several classification tasks that are studied in federated learning setting in healthcare, e.g., autism spectrum disorder (ASD) [18], cancer diagnosis [41,43,49], COVID-19 detection [40,44,48,54], human activity and emotion recognition [20,21,37,42,45], patient hospitalization prediction [52], patient mortality prediction [19,39,46,47,55,56], and sepsis disease diagnosis [51]. The summary of classification tasks in FL studies for medical application is listed in Table 5.…”

“…Cancer diagnosis. Recent studies show that researchers are employing FL technology to develop machine learning models for cancer diagnostic applications [41,43,49]. For instance, Lee et al [27] proposed a CNN-based model to classify whether thyroid nodules were benign or malign.…”