Convolutional Siamese networks for one-shot malaria parasite recognition in microscopic images

Malaria is a life-threatening, mosquito-borne disease caused by a Plasmodium parasite that is commonly spread to humans from the bite of Anopheles mosquitoes. It is widespread across the world, particularly in tropical regions, and if not treated in time, it can be deadly. Thus, early and effective treatment and diagnosis of malaria will save lives. Modern machine learning and image analysis methods are practiced on microscopic images of blood smears by using hand-crafted features that need expertise in interpreting these features. In contrast, convolutional neural networks (CNNs) have proved to be adequate in feature extraction and classification of parasite detection toward malaria diagnosis. CNNs' techniques can extract highly influenced features, determine filters to serve as an effective prediction that provides an accurate detection to aid malaria diagnosis. The motivation of the research is to state the significance of deep learning techniques in the diagnosis of malaria. This research proposes the convolution Siamese network (CSN) for a one-shot recognition method for parasite recognition in microscopic images. In this study, microscopic images were used to examine the performance of the CSN model, and it will assist in diagnosing various infectious diseases rapidly. CSN is implemented in a two-phase criteria to detect similarity or dissimilarity. The first phase consists of extracting detailed features using a fully connected convolutional block and the second phase consists of a similarity check where inputs are discriminated with similarity measures (L1, L2). This discriminated feature is fed forward to the final layer with a sigmoid function that depicts whether it is similar to the given input. Further, the proposed model is evaluated to attain optimum generalization, and its performance is captured by splitting data into identical divisions (50–50). As the model is trained only on half the malaria blood samples, it gives precisive generalization and attains malaria test accuracy scores of 87.10% and 87.38% with L1 and L2, respectively. We also depicted the mean square error (MSE) rate and confusion matrix for each to incorporate the performance for similar and dissimilar identification individually. The evolution of these metrics described the stability of the proposed network. Experiments and results have proven that the proposed CSN can leverage performance and build comprehensive models with fewer samples. The obtained result shows the potential of the Siamese network in the field of medical diagnostic tools. This research is the first study to construct and assess the Siamese network diagnostic model toward malaria diagnosis.