A Patient Specific Algorithm for Plasmodium Malaria Detection on Cell Images

Computer-aided detection of plasmodium malaria on cell images from digital microscopy provides an invaluable second opinion to medical experts. Traditionally, well-established deep learning algorithms are widely used to detect plasmodium, but these techniques have not been deployed due to the uninterpretable nature of the decisions made by the network. To address this, we present an ensemble method that fuses the predictions based on the class activation mapping (CAM) results from multiple networks via a novel selector network. We study the performance on a publicly available dataset of cell images by using three convolutional neural network architectures, Xception, ResNet50, and InceptionV3,to produce CAM results for the selector network and performance comparison. Our proposed approach helps enhance doctors' trust and outperforms the traditional methods by 1%.