A multi-center distributed learning approach for Parkinson's disease classification using the traveling model...

Area of Science:

• Medical Imaging
• Machine Learning
• Neurology

Background:

• Centralized machine learning (ML) faces challenges in healthcare due to data privacy concerns.
• Previous federated learning (FL) and traveling model (TM) studies used limited centers or simulations, questioning real-world applicability.
• The traveling model (TM) setup is effective for limited data but hasn't been applied to medical image classification.

Purpose of the Study:

• To develop and evaluate a convolutional neural network (CNN) for Parkinson's disease classification using the TM setup.
• To assess the effectiveness of TM in a real-world scenario with diverse data from numerous centers and limited samples.
• To demonstrate the applicability of TM for 3D medical image classification.

Main Methods:

• Utilized a convolutional neural network (CNN) architecture.
• Employed the traveling model (TM) distributed learning setup.
• Trained the model on 3D medical imaging data from 83 diverse real-world centers, with most contributing small training samples.

Main Results:

• The TM-trained CNN achieved an Area Under the Receiver Operating Characteristic Curve (AUROC) of 83%, matching or slightly surpassing the centrally trained model's 80% AUROC.
• Demonstrated the effectiveness of TM in complex, real-world distributed learning scenarios with heterogeneous data distributions.
• Showcased TM's capability in training CNNs even with limited training samples per center.

Conclusions:

• The traveling model (TM) is effective for 3D medical image classification, particularly with limited and heterogeneous distributed data.
• This approach is highly relevant for training ML models using data from small, remote medical centers, and for rare diseases.
• The simplicity of TM enables broad application to various deep learning tasks, enhancing clinical utility across diverse medical facilities.