Reducing data dimension boosts neural network-based stage-specific malaria detection

Although malaria has been known for more than 4 thousand years(1), it still imposes a global burden with approx. 240 million annual cases(2). Improvement in diagnostic techniques is a prerequisite for its global elimination. Despite its main limitations, being time-consuming and subjective, light microscopy on Giemsa-stained blood smears is still the gold-standard diagnostic method used worldwide. Autonomous computer assisted recognition of malaria infected red blood cells (RBCs) using neural networks (NNs) has the potential to overcome these deficiencies, if a fast, high-accuracy detection can be achieved using low computational power and limited sets of microscopy images for training the NN. Here, we report on a novel NN-based scheme that is capable of the high-speed classification of RBCs into four categories-healthy ones and three classes of infected ones according to the parasite age-with an accuracy as high as 98%. Importantly, we observe that a smart reduction of data dimension, using characteristic one-dimensional cross-sections of the RBC images, not only speeds up the classification but also significantly improves its performance with respect to the usual two-dimensional NN schemes. Via comparative studies on RBC images recorded by two additional techniques, fluorescence and atomic force microscopy, we demonstrate that our method is universally applicable for different types of microscopy images. This robustness against imaging platform-specific features is crucial for diagnostic applications. Our approach for the reduction of data dimension could be straightforwardly generalised for the classification of different parasites, cells and other types of objects.

Automated summary

Malaria, a global burden with millions of annual cases, requires improvement in diagnostic techniques for its elimination. This study introduces a novel neural network-based scheme capable of high-speed and accurate classification of malaria infected red blood cells. The smart reduction of data dimension significantly improves the performance and applicability of the method.