Ensemble Transfer Learning for Accurate Malaria Detection with Stage Classification & Explainable AI

Malaria is a mosquito-borne disease caused by the Plasmodium parasite and spreads amongst humans by the bite of the female Anopheles mosquitoes. Conventional detection of the disease relies on examining the travel history of the patient to malaria endemic countries, later blood smears are analyzed under a microscope to detect the presence of malaria. This diagnostic approach requires specialized expertise, a resource often scarce in malaria burdened countries. This paper proposes a novel method to address these issues, to not only classify the presence of malarial infection but also identify the current stage of the malarial infection based on available blood smears. The proposed methodology employs deep transfer learning from multiple pre-trained Convolutional Neural Network (CNN) models (VGG16, VGG19, MobileNetV2, Xception, InceptionV3 and a custom trained model). To provide an accurate representation of the best performing models the models have been given weights and the classification is done through majority voting for the classes. This method provides a tested solution for datasets that are unbalanced and for models that are unable to fit the dataset with high accuracy. To also understand how the model can make this prediction it uses Gradient-weighted Class Activation Mapping (Grad-CAM) to understand which parts of the image are important for the image classification. This method has resulted in a higher overall accuracy and reduced false negative rate. The model has provided a method to help malaria diagnosis and potentially aid in disease control and management.