Performance Evaluation of EfficientNet Model Towards Malaria Parasite Detection in Segmented Blood Cells from Thin-Blood Smear Images

Abstract

Malaria is a potentially fatal disease caused by infected Anopheles mosquitoes. Traditional diagnosis of malaria involves examination of thin blood smear slides under a microscope by trained microscopists to detect infected blood cells. This process is expensive, and results depend both on the quality of smear and on the expertise of the microscopist. Thus, work done in this field is focussed on automating this detection of infected cells. Early work for this task included using hand-engineered features and machine learning algorithms. This approach was taken over by the advent of CNNs which provided an end-to-end solution, right from feature extraction to classification. Work done in this field thus shifted towards using state-of-the-art CNNs. The authors of this paper found that most of the existing models that had been used for this problem had good classification accuracy but had big sizes that could not be run on a normal computing device. They identified that a new set of state-of-the-art CNNs, called EfficientNets which promised similar performance at much smaller sizes, had not been used for this task. So, they worked to evaluate the performance of EfficientNet models for this medical image classification task. The authors trained the EfficientNet-B0 model on the malaria images dataset taken from the National Library of Medicine (NLM) (Dataset used: https://lhncbc.nlm.nih.gov/LHC-publications/pubs/MalariaDatasets.html ). The authors then computed the key performance parameters of the model and compared them against the existing models for this problem. The work done showed that using Efficient-Net models for this task achieved similar accuracy to existing models and significantly reduced the number of parameters.