Supervised learning and image processing for efficient malaria detection

Malaria is a devastating disease that leads to many deaths each year. Currently, most malaria diagnoses are performed manually, which is time consuming. This may result in it taking longer to diagnose patients, especially those in poor and rural areas, motivating the development of automated detection tools. Deep learning approaches, particularly convolutional neural networks (CNNs), have seen success in the existing literature. However, CNNs are computationally expensive and require significant amounts of training data, which may limit their real-world viability, especially in poorer and rural communities. Nondeep supervised techniques are largely free from these limitations but have received less attention in the existing literature. This paper differs from existing work using non-deep systems by investigating the use of RFs, adopting a more rigourous testing methodology and conducting a broader exploration of pre-processing techniques. Two non-deep supervised systems are proposed, based on random forests (RFs) and support vector machines (SVMs). The RF system performs better, having achieved an accuracy of 96.29% when tested on 20 000 images, with runtimes of less than two seconds. Testing on a small dataset of images gathered from a different source achieves similar performance, suggesting the model may generalise to different imaging conditions. The system achieves higher recall than existing non-deep approaches, and its accuracy, recall and precision are within 4% of the highest performing CNN approach.