Malaria Detection using Open Microscope and Deep learning with Jetson nano

An open-source microscope that can detect disease like malaria, the main goal is to give quality health checkup to poor people.

We used Jetpack SDK with Python and Tensorflow to inference the data model also we found an article on medium that describes the Malaria Dataset: Source

A subset of the Malaria Dataset provided by the National Institute of Health (NIH). We will use this dataset to develop a deep learning classification model.

The malaria dataset we will be using in today’s deep learning and medical image analysis tutorial is the exact same dataset that Rajaraman et al. used in their 2018 publication.

The dataset consists of 27, 588 images belonging to two separate classes:

The project is fully open-source and We believe this project can make big changes in society by helping to give health and proper treatment to the poor. And with more data models and training, we can improve accuracy and better fight with diseases like Malaria. if we have a powerful GPU system we could train more datasets and provide models, will try that in future. Finally, we have a complete end-to-end malaria classification system, Thank You all.

Project Link:

1 Like


You are right about the need, and it would be nice to have such a low cost tool for malaria screening.

However, I don’t wanna cast a cloud, but from my own understanding of this topic you have some hard issues to overcome, so be aware of these for deciding if you invest more time:

  • from the image above, I see bad quality smear (many erythrocytes seem damaged on right side). For using a such a model, you would need to get similar image as the dataset. Which is not the case, yet.
  • In same image I see unfocused cells on the left. You would have to set this under control as you are looking for small details.
  • I also see non homogeneous lighting with some green/blue light in the middle and pinky color around. You would also need to have this under control.
  • Staining may not be so obvious. The colors depend on many physico-chemical factors including staining compounds, temperature, humidity, drying of the blood smear. It may also depend on if you use methanol or not, the former may be dangerous, but alternatives are still expensive.
  • You would have to classify red blood cells from nucleated blood cells and platelets (and platelets aggregates, many others…) in order to get RBC images only for the classifier. You may segment the image and use metrics on these to get RBCs only. Also you would have to deal with amount of RBCs. There may be many double or more RBCs stuck together, so you would have to perform a search for a part of the slide with convenient RBC density.
  • I don’t know the Malaria dataset, but be aware that many other RBC inclusion or abnormal forms may be found, from genetic diseases such as hemoglobinopathies (that have been selected by malaria itself, making people with abnormal hemoglobin to be less affected, so these are more frequent in endemic regions with higher malaria prevalence). Other pathologies or parasites may also interfere.
  • Again I don’t know the dataset, but the signal would be higher for mature forms (schizonts) of the parasite than for immature forms such as rings.
  • There are different plasmodium species. Plasmodium falciparum is the more dangerous, but mature forms of this plasmodium are not frequent in venous blood because it adheres to capillary vessels in order to escape spleen entrapment. So for this case you would need a very good resolution as well.

Automated smearing and staining robots would be far too expensive, but I have seen these last years a manual mechanical smearer device that you may consider for reproducible smear (this one was called HemaPrep).

You would also have to consider staining with environment and drying for reproducible image.
As far as I know, the leaders in digital microscopy for hematology still have problems with slight variations in smearing/staining. Although the human eye+brain easily accomodates, machine algorithm get more issues with this. You may try to add this in data augmentation.

As far as I know, a leader in digital microscopy, although using automated smearing/staining and micrometric plate + high quality microscope digitalization has proven very poor results in malaria detection (especially for Pl. falciparum) in its Advanced RBC application.

  • Typical malaria infestation is about 1 infected RBC per 1000 RBCs. Ideally a detection threshold of 1/15000 RBCs would be wanted, but let’s admit 1/5000. You need to confront this with the 97% accuracy of the NN classification. This would not allow a reliable detection unless in very high parasitemia.

Finally, your system would require, per sample, a glass slide with coating, a smearing/staining/drying in a device with controled conditions, images acquisition with controled lightening and focusing, some more segmentation/classification models. Be aware that searching on AliBaba for antibodies or better antigenic Malaria Rapid Diagnostic Tests (RDT) you can find many with pan malaria band and falciparum specific band below 1 $.

Anyway, some say the most important is not the goal but the travel itself. Porting the model to the Nano will probably be the easiest and funniest part. For other issues, you may have to be patient, but I’d be sincerly happy if you show that I was wrong and if you find new solutions.

Feel free to contact me by PM for further advice. Good luck.


That image might also need a non-lossy image format. I don’t know if it is just from posting to the forum or not…perhaps the original was a non-lossy format…but the image as posted seems to use a lossy compression (JPEG is bad…I like PNG).

1 Like

Thank you for your valuable feedback, Highly appreciate that ,and very excited to get these much parameters to look, I’ll check each one and get back you.

1 Like

Thanks for the feedback, you are right , the image in lossy format, that because I was using 3D Printed DIY Microscope and modified raspberry cam , I was planning to use more stable microscope and better camera to get the cell image.

1 Like