A transfer learning approach to identify Plasmodium in microscopic images.

Area of Science:

• Medical imaging and diagnostics
• Computational biology and bioinformatics
• Infectious disease research

Background:

• Malaria, caused by Plasmodium parasites, is a major global health issue, responsible for millions of cases and hundreds of thousands of deaths annually.
• Accurate identification of malaria in blood smears typically requires specialized expertise, posing a challenge for widespread clinical application.
• Existing automated methods for malaria detection need further refinement for reliable clinical use in Computer-aided Diagnosis (CAD) systems.

Purpose of the Study:

• To evaluate the effectiveness of transfer learning using pre-trained deep learning architectures for the automated identification and classification of malaria.
• To assess the performance of the DenseNet201 model in detecting Plasmodium parasites within microscopic blood images.

Main Methods:

• A dataset comprising 6222 Regions of Interest (ROIs) was utilized, including images from the Broad Bioimage Benchmark Collection (BBBC) and locally acquired images from Brazil.
• The dataset underwent rigorous cross-validation using 100 distinct partitions (80% training, 20% testing) with circular ROIs.
• Transfer learning was applied using well-established deep learning architectures, with a focus on DenseNet201 for performance evaluation.

Main Results:

• The DenseNet201 model achieved a high Area Under the Curve (AUC) of 99.41% for identifying Plasmodium parasites in ROIs, indicating excellent diagnostic performance.
• The model demonstrated fast processing times, making it suitable for rapid analysis.
• DenseNet201 significantly outperformed other considered networks, with a 99% confidence interval, highlighting its superior capability in malaria detection.

Conclusions:

• Transfer learning, particularly with the DenseNet201 architecture and texture features, shows significant potential for differentiating malaria-infected individuals.
• The approach is capable of detecting Plasmodium parasites even within challenging leukocyte images.
• Future work will focus on scaling the approach with more data and developing a user-friendly interface for clinical CAD implementation, aiming to benefit global populations, especially those near the Amazon region.