HarmonyTM: multi-center data harmonization applied to distributed learning for Parkinson's disease classification.

Area of Science:

• Medical Imaging
• Machine Learning
• Distributed Learning

Background:

• Distributed learning enables training machine learning (ML) models on diverse datasets while adhering to data-sharing regulations.
• The traveling model (TM) approach is beneficial for limited local datasets but is susceptible to scanner-induced biases when centers use different acquisition devices.
• Existing data harmonization methods often require large or paired datasets, which are impractical in distributed settings.

Purpose of the Study:

• To introduce HarmonyTM, a novel data harmonization technique specifically designed for the TM approach in distributed learning.
• To mitigate feature representation bias caused by multi-center scanner variability in neuroimaging datasets.
• To improve the accuracy of ML models in clinical applications, such as Parkinson's disease classification, by preventing reliance on scanner-specific artifacts.

Main Methods:

• HarmonyTM employs adversarial training to remove scanner-specific biases from features used in ML models.
• The method is tailored for the traveling model (TM) framework, enabling sequential training across multiple centers.
• Evaluation was performed on multi-center 3D neuroimaging datasets from 83 centers utilizing 23 distinct scanners.

Main Results:

• HarmonyTM enhanced Parkinson's disease (PD) classification accuracy from 72% to 76% within the TM setup.
• The method significantly reduced the model's ability to classify data based on scanner type, decreasing accuracy from 53% to 30%.
• These improvements were achieved without the need for large or paired datasets.

Conclusions:

• HarmonyTM effectively harmonizes 3D neuroimaging data in distributed learning settings using the TM approach.
• The method successfully minimizes shortcut learning by preventing classifiers from exploiting scanner-specific variations.
• HarmonyTM is crucial for developing robust and clinically applicable ML models by ensuring disease classification is independent of data acquisition hardware.