A Privacy-Preserving Approach to Effectively Utilize Distributed Data for Malaria Image Detection.

Area of Science:

• Medical imaging analysis
• Machine learning in healthcare
• Parasitology

Background:

• Malaria, caused by Plasmodium falciparum, infects red blood cells, necessitating early detection systems.
• Visual complexity in malaria datasets requires extensive data for effective machine learning (ML) model training.
• Data sharing is hindered by privacy regulations like GDPR and DPA, limiting collaborative research.

Purpose of the Study:

• To develop an effective computer-aided system for early malaria detection.
• To overcome data sharing limitations using federated learning (FL).
• To evaluate the performance of ML models within an FL framework for malaria image analysis.

Main Methods:

• Utilized a federated learning (FL) framework with real-time medical image data.
• Employed state-of-the-art ML models: DenseNet and ResNet-50.
• Experimented on a malaria dataset of 27,560 images across different client configurations.

Main Results:

• DenseNet achieved higher accuracy (75%) than ResNet-50 (72%) with eight clients.
• Both models showed similar high accuracy (94%) with four clients.
• DenseNet (92%) outperformed ResNet-50 (72%) with six clients.

Conclusions:

• Federated learning enhances ML model accuracy for malaria detection through decentralized learning and local adaptation.
• The FL framework effectively ensures data privacy, complying with GDPR and DPA.
• This research demonstrates the viability of FL for collaborative medical image analysis while preserving patient data confidentiality.