Morphology classification of malaria infected red blood cells using deep learning techniques

Malaria is an endemic disease that causes great harm to children and pregnant women. Without early and proper diagnosis, it leads to organ failure, coma and eventually death. The gold standard technique of diagnosing malaria is the thick and thin blood smear microscopy which entails the visual inspection of a blood smear slide under a microscope for detecting malaria parasites which inhibit the red blood cells (RBC). This technique is highly subjective, tedious and time consuming, it also requires expert skill in malaria microscopy which is highly lacking in malaria endemic regions. To tackle these drawbacks, a lot of studies have automated the process using Artificial Intelligence, but these systems focus on the detection of malaria parasites only. Clinical microscopy for diagnosing malaria goes beyond detecting parasites in blood smears, other abnormalities seen by the microscopists are recorded as well. One such common abnormality is the rouleaux formation, which is the stacking of red blood cells like chains of coins, its presence indicates the presence of an infection. To develop truly automated systems capable of being deployed in low resource settings, these systems need to be familiar with this highly common occurring deformity of the RBC morphology (Rouleaux formation). Hence this study developed a dataset of 12,356 750x750 pixel images of Rouleaux formation morphology and 12,356 750x750 pixel images of normal RBC morphology. Five different CNN architectures were trained and tested to benchmark the dataset for the binary classification of the dataset dataset with DenseNet121 achieving the highest accuracy of 99%.