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

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Machine Learning

Background:

• Federated learning (FL) is a decentralized machine learning strategy that adheres to medical data privacy regulations.
• FL enables the generalization of deep learning algorithms by sharing only models and parameters, not raw patient data.
• This study applied FL to ultrasound image analysis for predicting benign or malignant thyroid nodules.

Purpose of the Study:

• To evaluate the performance of federated learning (FL) in thyroid nodule classification.
• To compare the efficacy of FL against conventional deep learning (DL) methods.
• To assess the potential of FL in medical image analysis while preserving patient privacy.

Main Methods:

• Utilized a dataset of 8457 ultrasound images (5375 malignant, 3082 benign) from six institutions.
• Employed five deep learning networks: VGG19, ResNet50, ResNext50, SE-ResNet50, and SE-ResNext50.
• Performed internal validation on 20% of the data and external validation on 100 images from an independent institution.

Main Results:

• Internal validation showed Area Under the Receiver Operating Characteristic (AUROC) curves for FL ranging from 78.88% to 87.56%.
• Conventional DL achieved AUROC curves between 82.61% and 91.57% for internal validation.
• External validation yielded AUROC curves of 75.20%–86.72% for FL and 73.04%–91.04% for conventional DL.

Conclusions:

• Federated learning performance is comparable to conventional deep learning when analyzing decentralized medical data.
• FL offers a privacy-preserving alternative for medical image analysis.
• This approach holds potential for widespread adoption in clinical settings requiring data confidentiality.