I. Introduction

Collaborative Disease Diagnostics (CDDx) is a computer vision and machine learning company in the medical diagnostics market. We are developing image pattern recognition software to be used as a smartphone application for malaria microscopy diagnosis. Malaria is one of the leading causes of death in developing countries, affecting children under five and pregnant women in particular. After consulting Dr. Ian Crandall, a malaria researcher at the University of Toronto, we have validated that manual microscopy diagnosis of malaria is highly prone to human error, which negatively impacts costs and the health of the population. Automating microscopy diagnosis through software designed by CDDx will eliminate these issues.

CDDx’s custom algorithm will be designed to have 95% accuracy in malaria diagnosis, on par with the current gold standard for diagnostic tools. Alternative methods of diagnosis include the use of Rapid Diagnostic Tests to quickly detect the presence of malaria. Unfortunately, these tests have a limited shelf life that is influenced by environmental temperature and humidity. Other portable diagnostic tools require disposable cartridges to be used in conjunction with their products. Unlike that which is offered by CDDx, these competitors do not present cost-effective and sustainable solutions. We hope to apply our image recognition software in a smartphone application, so when used in conjunction with lens attachments, on-site images of malaria blood smears can be analyzed to assist diagnosis in a timely manner.

II. Approach

Using technology to automate malaria diagnosis via smartphone is very efficient, economical, and accurate. The core component of our product incorporates a combination of computer vision and machine learning techniques. Once this foundation is built, minimal infrastructure will be required to maintain and use our product, providing a sustainable and cost-effective solution. As of date, we have successfully completed the cell counting and red blood cell identification module of the application. Our next step is to accurately identify the malaria parasite from other elements present on blood smear slides viewed through the microscope attachment. We are currently utilizing various image capture and image processing methods to extract data from the images that are characteristic of malaria. The success of our classifier is dependent on the success of this step, as it is critical that malaria is identified with a 95% accuracy rate.

Our architecture consists of four components: image processing and rendering, detection, feature extraction, and classification. We are currently focusing our efforts on creating a proof of concept for doing cell detection.

The image manipulation to do the blob detection consists of three components, which, in turn, produce an image that can be used to detect cells. First, JPG images of blood smears are taken and converted into black and white images. We are currently interested in image contrast rather than image colour, as the differences in pixel intensity will be used to determine the edges of cells. The next step is image inversion, which is necessary for the success of the subsequent tool, erosion, which works by making lighter objects smaller than the darker objects. The point of using erosion is to try to make the cells more distinct from each other. From here we branch off by using two different techniques: one for the image detection and the other for determining overlapping cells.

The first technique, difference of Gaussians, is used to do blob detection of cells. This is done by subtracting a blurred version of the image from the original image, which enhances the edges of the original. Next, the watershed segmentation technique is implemented to detect overlapping cells. By determining the local minimums of the image, this method is able to segment the original image into its constituent sub-differentiated objects. This is incredibly important for identifying red blood cells since these cells tend to overlap – especially on slides that have not been well-prepared.

Once the cell detection is done properly, we can develop an algorithm for identifying cells infected by malaria parasites. At this stage, we will be looking for several key characteristics on the images captured to determine malaria-infected cells. The first is chromatin, which is a part of the cell nucleus. It encircles the cell, and stains bright red. Next is the cell cytoplasm that stains blue once the Giemsa dye is absorbed by the cell. Additionally, pigment is a granular by-product of parasite growth, but it does not take up stain, hence a golden-brown to black colour. Finally, there is the presence of Shuffner’s stippling dots in the cells. Schuffner stippling are masses of pink dots that commonly appear in some P. vivax parasitized red blood cells.

In order to achieve parasite identification, the algorithm will use a similar method to the one previously stated. The difference is that we must use adaptive threshold in the watershed segmentation technique to localize malaria parasites. The algorithm will determine if the parasites are within a particular cell using the boundary-drawing algorithm. Using these techniques allows us to accurately count the number of cells, find infected cells, and then crop the images for feature extraction.

III. Conclusion

By improving the accuracy of diagnosis in the absence of trained healthcare professionals, CDDx offers a valuable product that has the capacity to save human lives. With faster, accurate diagnosis, healthcare practitioners will be able to offer prompt treatment to decrease deaths by malaria. Our initial goal is to automate identification of malaria-infected cells, but we hope to eventually be able to distinguish different strains of malaria. CDDx aims to launch a Minimum Viable Product by early February 2016. As a computer vision and machine learning company, CDDx hopes to expand to identifying other infectious diseases in the future.