Malaria Prediction using Fusion Learning with Enhanced Trust and Transparency
Anjali Gangone, Abiy Abraham, Swapna Gangone, Bharat Kumar G J · 2024
Malaria is a dangerous infectious disease that kills a lot of people globally every day. It is typically diagnosed by trained personnel visually inspecting blood smears under a microscope for parasite-infected red blood cells. This method is ineffective, and the diagnosis is based solely on the examiner’s knowledge and experience. Malaria blood smears have previously been subjected to machine learning-based automatic image recognition techniques for diagnosis. However, performance in practice has been inadequate. As a result, for malaria prediction, the study used deep learning. With enhanced trust and transparency, the model makes decisions and it provides explanations for model predictions in a human-readable format. In addition, the study describes in detail dataset preprocessing, feature extraction, dataset preparation, development tools, and several assessment criteria. In turn, the VGG-19, ResNet-50, Inception V3, and EfficientNetB4 of Conventional Neural Network algorithm with transfer learning classified malaria as parasitized or uninfected by using Anaconda software. The first set of information was used from the online medical records of patients who fit the standard profile. Then, the study used 80% for training, and 10% for validation while the remaining 10% was used for testing within blood smear images the size of 224*224 pixels. In terms of model performance, ResNet50 has the best classification performance with an accuracy value of 94.37%; than EfficientNetB4, In-ceptionV3, and VGG-19. The Implement model used interpretability techniques LIME (Local Interpretable Model-agnostic Explanations) to produce human readable explanations to enhance trust and transparency.