Simulating Federated Transfer Learning for Lung Segmentation using Modified UNet Model.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• The COVID-19 pandemic underscored the need for rapid AI-driven lung disease diagnosis from X-rays.
• Data privacy concerns and the requirement for large, centralized datasets hinder AI model development.
• Federated Learning (FL) offers a decentralized approach to train AI models without compromising data privacy.

Purpose of the Study:

• To simulate a Federated Transfer Learning framework for lung segmentation on X-ray images.
• To address data privacy challenges in AI model training for medical diagnostics.
• To improve the accuracy and efficiency of lung segmentation in clinical settings.

Main Methods:

• Simulated Federated Transfer Learning (FTL) using a U-net model pre-trained on MRI data.
• Leveraged local healthcare data at each node to augment training datasets.
• Employed Explainable AI (XAI) techniques to interpret model predictions.

Main Results:

• Achieved near-perfect lung segmentation accuracy using the proposed FTL approach.
• Demonstrated improved model performance by utilizing similar local healthcare data.
• Validated the effectiveness of FTL in handling distributed datasets of varying sizes.

Conclusions:

• Federated Transfer Learning is a viable solution for privacy-preserving AI in medical image analysis.
• The integration of local data and pre-trained models enhances segmentation accuracy.
• This approach offers a scalable and secure method for developing diagnostic AI tools.