Fighting Malaria with Technology
Posted on April 24, 2015
Malaria is a disease caused by a protozoa and transmitted by a mosquito. Estimates suggest it is the cause of 500 thousands deaths annually. The World Health Organization estimates that almost 3 billion people are at risk of Malaria and cited several challenges in their latest Malaria report for 2014. Among these challenges were the diagnostic tests deficiencies.
There are currently 4 approaches to diagnose Malaria: Clinical symtoms/history, Blood tests using microscopy, Rapid tests using a Dip stick and Molecular level tests. There are pros and cons for each of these with the dip stick test becoming popular for its relative speed and cost attributes.
We wanted to do something to help.We are interested in the ‘real world’ and how technology can and should make a difference for real problems in the physical world. We want to do things that will move the needle, can bring an order of magnitude or better improvements and contribute to humanity in a big way. This problem, given the mortality rate, the population at risk really got to us and it fit perfectly with our passion for helping people and modernizing health care with technology.
We thought about the diagnosis challenge and asked ourselves some questions:
What happens if a whole village needs to be tested?how long will it take?how accurate the results would be?Would this data, should it be digitized, improve our understanding of Malaria? What risks are there for the medical and support staff in the field while doing this?
There were indeed many questions, and we came to the conclusion that in a crisis situation when there is more than a handful of people to diagnose and more generally when dealing with epidemics there must be severe challenge in carrying out the necessary tasks at all, let alone efficiently.
We decided that our contribution will be in using our machine learning and computer vision expertise to build a comprehensive system that can deal with both clinical symptoms and blood sample image. We envisioned that we will have a 2-steps process where clinical symptoms will signal whether a blood sample should be taken and the system can improve overtime through feedback and fusion of the data from the different steps.
We started collecting the necessary data including cases that had similar symptoms but actually negative diagnosis. For example sometimes a person presents with symptoms that look like Malaria but turns out to be Ebola. We also constructed a dataset of blood sample imagery for both healthy and infected cells. The images were in PNG formats but could have been DCIM or other common formats used in microscopy. The dataset construction is probably one of the hardest, if not the hardest, part. We needed to apply natural language algorithms to normalize the clinical symptoms. We also needed to apply a set of transformations in order to prepare the data for our algorithms. We evaluated different machine learning models including neural and geometrical machine learning models and found the geometrical models with a kernel transformation to be more reliable. We decided that a non-linear kernel model in a low order polynomial was the best choice. This model was 100% accurate in the positive case and 80% accurate in the negative case.What this means is that 2 times out of 10 we may draw a blood sample unnecessarily, which we believe to be an acceptable cost. We need to expose the system to more data in order to verify the performance over a large sample. More details will be available in a paper to be published later this year.
In parallel, We worked on the blood sample images. We evaluated different approaches given what we had built so far already for our platform’s sensory system. This problem is much more specialized. The quality, the resolution and color gradients varied greatly. We thought that we can go even further than our initial goals of “just” classifying the infection: We can make it faster, better and easier to train new technicians and test them and we can enable collaboration between machines and humans when it comes to microscopy images. We built a component that can be used to view and label the images semi-automatically and can be used by universities and remote professionals to collaborate remotely.
We then focused on the classification of the cells. Based on our learning from the cell detection and segmentation phase any ‘never before seen’ image is segmented to detect the cells. Our system then assigns each cell a class of ‘Infected/Not-infected’. This means we can calculate all kind of interesting statistics like infection rate for example.
You can see the image below, the rectangles are automatically drawn to note which cells are infected.
Both of the diagnosis support system, clinical or visual, run within 100s of milliseconds. This is the kind of impact we were looking to make and we believe it is the kind of impact that is needed to address large outbreaks or , worse, epidemics.
How can this work in the field?
There are 3 distinct use cases that are excellent applications for CellMates:
- University/Training: Using our technology, we can train and assess more professionals faster.
- In-lab Microscopy: We believe an IoT Microscope would be a great use case of our platform.
- Mobile Microscopy: Mobile microscopy is low cost, portable devices that can be plugged into networks or tablets. Relatively cheap lightweight microscopes already exist and We believe that in general the optics to support even smaller and more economic mobile microscope are also available.This is the use case that can make the difference and tip the scale in epidemics and usher in a new way to response especially in high-risk of contagious infections.
Deploying and Using CellMates:
CellMates is available for Linux based systems, including Android. This makes it ready to integrate with any of the available mobile microscopy solutions. It works both online and offline meaning it can be taken into rural villages to support the professionals in the field.
We are offering CellMates free of charge to under funded non-profits, and very poor nations. We would be honored and delighted to support these teams in their fight for life.
For other types of users, a flat subscription is possible with unlimited installations as well as per embedded device licensing.
Our CellMates’ work is now integrated into our platform and is available through our APIs. We are looking to partner with organizations focused on Malaria and other diseases. We are also looking to work with optics and microscope manufactures to realize a complete solution that can make a difference.
Malaria has 5 different sub-types, and we would our system to be even more specific in the diagnosis. For that we need to get more data points and images.
We have already started our fight against Epilepsy and Cancer. Stay tuned for good news.
Want to help us?
You can help us in different ways, including providing data and other means of support. You can help us by promoting and morally supporting this work which will hopefully inspire more technology companies to work on harder and more impactful problems.
To discuss more, please write to us at firstname.lastname@example.org.
There are a lot of challenges that we as human must face and surmount. Our planet, our societies and our individual wellbeing are facing new kinds of stress tests. Some of these challenges are really hard but they are the ones that can tip the scale and make the difference for generations to come. Even small startups can contribute to this end.Let ‘workarounds’ not be our goals and instead let us try to solve these grand challenges. We are committed to continuously expand the breadth and depth of our platform’s sensory and reasoning capabilities to help people live better and happier.
You can watch the video we made about CellMates here: