A Transformer-based Algorithm for Automatically Diagnosing Malaria Parasite in Thin Blood Smear Images Using MobileViT

Based on the report provided by the World Health Organization (WHO), malaria has proved to be a life-threatening disease whose cases reached 241 million in 2020 globally. However, diagnosing malaria in the early stages of infection can be very fruitful for ameliorating this disease. The standard way of diagnosing malaria is by examining the blood cell images by professionals. Despite medical technology development, this is not feasible in many underdeveloped areas due to the lack of such experts. Thus, researchers interested in computer-aided decision-making, specifically deep learning, have focused on atomizing the diagnosis of malaria recently. The performance of transformer-based models combined with convolutional neural networks motivated us to propose an approach based on MobileViT for atomizing the process of diagnosing malaria. To achieve this, the model was trained on blood cell images collected from a publicly available dataset. Evaluated on 27,560 samples, the proposed classifier achieves an accuracy of 98.37% on average using 10-fold cross-validation. Among 2756 test samples, the model achieves 34 false negatives at least and 48 ones at most. Due to the medical nature of our problem, this is significant because the model’s miss-cases of actual positive malaria-infected samples are low, making the accuracy and recall of the model 98.37% and 98.38%, respectively. To our knowledge, this is the first study that applies a transformer-based model to a problem with superior performance. In addition, it is a lightweight and mobile-friendly neural network which can be utilized in mobile applications.