CT-Malaria Detection via Adaptive-Weighted Deep Learning Models.

Area of Science:

• Medical Diagnostics
• Computational Biology
• Machine Learning

Background:

• Malaria diagnosis via thin blood smears is challenging in low-resource settings due to image variability.
• Inconsistent image quality can lead to critical false negatives, impacting patient outcomes.
• Developing a robust and reproducible diagnostic process is essential.

Purpose of the Study:

• To create a reliable and replicable pipeline for malaria smear image analysis.
• To integrate classical machine learning with deep learning for enhanced diagnostic accuracy.
• To improve reliability through data-driven ensemble methods.

Main Methods:

• A two-track approach combining real-time augmentation, feature extraction, and classical classifiers.
• Training end-to-end deep learning convolutional networks.
• Utilizing pairwise ensembling with optimized, data-driven weight selection.

Main Results:

• The two-track architecture demonstrated consistent accuracy improvements over baseline methods.
• Weighted ensembling significantly enhanced diagnostic performance and reduced variance.
• Optimized fusion minimized false negatives from subtle parasites and false positives from artifacts, achieving 96.35% accuracy.

Conclusions:

• Integrating augmentation, multiple modeling tracks, and optimal ensembling maximizes malaria smear classification accuracy.
• This approach offers a robust framework for improving malaria diagnostics.
• Further enhancements are possible through supplementary models and multi-class extensions.