Deep Machine Learning Model Trade-Offs for Malaria Elimination in Resource-Constrained Locations

Area of Science:

• Medical technology
• Computer science
• Public health

Background:

• Deep machine learning (DML) models are increasingly used in healthcare.
• Limited computational resources and energy availability in developing regions pose challenges for DML deployment.
• Previous successes in malaria elimination are threatened by factors like the COVID-19 pandemic.

Purpose of the Study:

• To evaluate computational and predictive performance trade-offs of DML models for rapid malaria case finding.
• To identify DML models that maximize malaria detection accuracy while minimizing resource and energy consumption.
• To facilitate the deployment of effective DML solutions in resource-constrained healthcare settings.

Main Methods:

• Experimental evaluation of four candidate deep learning models using a blood smear malaria test dataset.
• Comparison of models based on detection accuracy, memory usage, and inference time.
• Quantization techniques were applied to optimize model performance for mobile deployment.

Main Results:

• Quantized Basic Convolutional Neural Network (B-CNN) and MobileNetV2 models demonstrated superior malaria detection performance, achieving up to 99% recall.
• These models exhibited significantly lower memory usage (2MB for 8-bit quantized models) and shorter inference times (33-95 microseconds on mobile phones) compared to VGG-19.
• MobileNetV2 was selected for implementation in a mobile application due to its even lower memory footprint than B-CNN.

Conclusions:

• Optimized DML models like MobileNetV2 offer a viable solution for rapid malaria case finding in resource-limited environments.
• The developed mobile application can enhance malaria detection capabilities and support global elimination initiatives.
• This research provides a pathway to overcome challenges in deploying advanced AI for public health in developing regions.