Automated diagnosis of malaria in thick blood smear films: deep neural network approach
Malaria, a mosquito-borne life-threatening disease, is one of the major health hazards in Sub-Saharan Africa. The potency of the disease is exacerbated by delayed and inaccurate diagnosis using the standard microscopy method which requires a lab technician to observe the malarial parasites under a microscope. In this mode, observations of the diagnosis vary among technicians as they are judgmental and based on intuition. This variance in results compromises the accuracy of the microscopy method and inadvertently leads to cases of drug resistance in patients and high mortality rates. To improve disease control, accurate and timely diagnosis of malaria is required for prompt interventions. While conventional machine learning techniques were found to offer reasonable malaria parasite detection and parasitaemia determination, the solutions are not yet able to sufficiently address the limitations of the automated malaria diagnosis. Conventional machine earning is limited by the need for skilled manual engineering of features which is still a brittle and subjective procedure. This study proposes the development of practical Deep Neural Network (DNN) approaches for the automation of microscopic diagnosis of malaria in thick blood smear slide images in highly endemic and low resourced setting. DNNs are a class of machine learning that have capability to learn from input data automatically without the need for manual hand engineering of features. This approach coupled with transfer learning will shape the needed solution for microscopy imaging especially under constraints of small datasets and limited computational resources. This approach, coupled with transfer learning, will form the needed solution for microscopy imaging especially where there are limitations on small datasets and few computational resources. To this end, the work contributed to an end-to-end solution in the field of microscopy image analysis by proposing a novel classification and object detection method based on DNNs for malaria diagnosis in thick blood smear images. First, we implemented an imaging apparatus with a 3D printable adapter which enables the attachment of smartphones to the eyepiece of a microscope. Using the developed apparatus, we collected and assessed the quality of thick blood smear film datasets for our machine learning tasks. Secondly, an automated Convolutional-Neural-Network-based classification model for the differentiation of malaria positive patches (parts of an image with parasites) from the negative patches (parts of an image without parasites) was developed. Secondly, a CNN-based classification model for the differentiation of malaria positive patches (parasites) from the negative patches (parts of an image without parasites) was developed. The solution has a wide application to not only malaria parasite detection but also for many microscopy health challenges like tuberculosis bacilli and intestinal parasites identification. Thirdly, the research also undertook investigations on multi-class detection task of trophozoites and WBCs identification and localization. In addition, an automated count of trophozoites and WBCs was generated from the detections, which aids in the calculation of malaria parasitemia, the quantitative measure of malaria infection in thick blood smear image dataset. Finally, mobile phone and web-based malaria screening systems based on the developed deep neural networks were developed. Such remote screening systems have enormous potential to provide quick and reliable microscopy diagnosis advancing microscopy disease pathology field especially in highly endemic but low resourced settings.