A comparative analysis of deep learning architectures on high variation malaria parasite classification dataset

Area of Science:

• Medical Diagnostics
• Artificial Intelligence in Medicine
• Parasitology

Background:

• Malaria diagnosis relies heavily on microscopy, facing challenges due to subtle visual differences between infected and uninfected cells.
• Computer-aided diagnosis of malaria is complex due to fine-grained variability in cell appearance.

Purpose of the Study:

• To create the largest malaria cell classification dataset to date (63,645 cells).
• To evaluate state-of-the-art deep neural network (DNN) architectures for malaria cell classification.
• To assess the impact of synthetic image generation on addressing class imbalance in malaria diagnosis.

Main Methods:

• Transformed an object detection dataset into a classification dataset for malaria cells.
• Evaluated multiple DNN architectures pre-trained on natural and medical images.
• Conducted qualitative analysis and evaluated models on an independent test set.
• Investigated the effect of conditional image synthesis on malaria parasite detection.

Main Results:

• Established a new benchmark dataset for malaria cell classification.
• Identified high-performing DNN models for malaria diagnosis.
• Demonstrated the generalizability of models on an independent test set.
• Showcased the utility of synthetic images in mitigating class imbalance.

Conclusions:

• Deep learning models show promise for accurate malaria diagnosis using a large, curated dataset.
• Conditional image synthesis can effectively address class imbalance issues in malaria detection.
• The developed dataset and model evaluations provide a foundation for AI-driven malaria diagnostic tools.