Build an image classification model that predicts the presence of leukemia in blood panels to assist with early detection and diagnosis.
Last updated
Was this helpful?
Created By:
Public Project Link:
GitHub Repo:
Acute Lymphoblastic Leukemia (ALL), also known as acute lymphocytic leukemia, is a cancer that affects the lymphoid blood cell lineage. It is the most common leukemia in children, and it accounts for 10-20% of acute leukemias in adults. The prognosis for both adult and especially childhood ALL has improved substantially since the 1970s. The 5-year survival is approximately 95% in children. In adults, the 5-year survival varies between 25% and 75%, with more favorable results in younger than in older patients.
Since 2018 I have worked on numerous projects exploring the use of AI for medical diagnostics, in particular, leukemia. In 2018 my grandfather was diagnosed as terminal with Acute Myeloid leukemia one month after an all clear blood test completely missed the disease. I was convinced that there must have been signs of the disease that were missed in the blood test, and began a research project with the goals of utilizing Artificial Intelligence to solve early detection of leukemia. The project grew to a non-profit association in Spain and is now a UK community interest company.
One of the objectives of our mission is to experiment with different types of AI, different frameworks/programming languages, and hardware. This project aims to show researchers the potential of the Edge Impulse platform and the NVIDIA Jetson Nano to quickly create and deploy prototypes for medical diagnosis research.
Pre-B Lymphoblastic Leukemia, or precursor B-Lymphoblastic leukemia, is a very aggressive type of leukemia where there are too many B-cell lymphoblasts in the bone marrow and blood. B-cell lymphoblasts are immature white blood cells that have not formed correctly. The expressions ("early pre-b", "pre-b" and "pro-b") are related to the differentiation of B-cells. We can distinguish the different phases based on different cell markers expression, although this is complex because the "normal profile" may be altered in malignant cells.
We had quite a lot of issues with the early class, this will be due to the fact the features of the different classes are very similar, especially so with the early class and benign as the early class represents the cells at the earliest stage of growth and is very close to the benign class.
From the project selection/creation page you can create a new project.
Enter a project name, select Developer or Enterprise and click Create new project.
Once downloaded head over the to Data acquisition in Edge Impulse Studio, click on the Add data button and then Upload data.
From here, choose 500 images from each class, ensuring you enter the label for the relevant class each time. The classes are not balanced so select around 500 images from each class, make sure to remember which images you have not used as we will use some to classify on device later.
By the time you have finished uploading you should have 2000 images uploaded into 4 classes, with an 80% / 20% train/test split.
The next step is to create your Impulse. Head to the Create Impulse tab. Next click Add processing block and select Image.
Now click Add learning block and select Classification then click Save impulse.
Head over to the Images tab and change the color depth to grayscale, then click on the Save parameters button to save the parameters.
If you are not automatically redirected to the Generate features tab, click on the Image tab and then click on Generate features and finally click on the Generate features button to generate the features.
The next step is to train our model. Click on the Classifier tab then change the number of training cycles to 150 and add an additional 2D conv/pool layer with 8 filter, 3 kernel size, and 1 layer to the network as shown in the image above. Now you are ready to start training, click the Start training.
Once training has completed, you will see the results displayed at the bottom of the page. Here we see that the model achieved 93.8% accuracy. Lets test our model and see how it works on our test data.
Let's see how the model performs on unseen data. Head over to the Model testing tab where you will see all of the unseen test data available. Click on the Classify all and sit back as we test our model.
You will see the output of the testing in the output window, and once testing is complete you will see the results. In our case we can see that we have achieved 91.67% accuracy on the unseen data.
Now we are ready to set up our Jetson Nano project.
On your Jetson Nano, run the following command and connect to the platform:
Once the script has finished running you will see a similar output to below:
For this project you will need opencv-python>=4.5.1.48. Make sure you have installed it on your Jetson Nano.
Now you can clone this project's GitHub repository to your Jetson Nano. Navigate to the location you would like to clone the repo to and use the following commands:
Now move the model you downloaded earlier into the model directory of the cloned repo, modifying your path as required:
Now it is time to add some test data to your device, ensuring that you are not using any data that has already been uploaded to the Edge Impulse platform. You can use additional images from the Kaggle dataset, or your own data. Place the images into the data folder of the cloned repo on your Jetson Nano.
The code has been provided for you in the classifier.py file. To run the classifier use the following command:
You should see similar to the following output. In our case, our model performed exceptionally well at classifying the various stages of leukemia, only classifying 3 samples out of 14 incorrectly.
In this demonstration, we have illustrated the potential of combining the Edge Impulse platform with the computational capabilities of the NVIDIA Jetson for swiftly prototyping medical research applications, such as early detection of leukemia.
Although this program is not yet capable of addressing the challenges associated with early detection or serving as a practical real-world solution, it serves as a valuable tool for medical and AI researchers, offering insights and aiding in the advancement of their respective fields.
NVIDIA Jetson Nano
Edge Impulse
For this project we are going to use the . Acute Lymphoblastic Leukemia can be either T-lineage, or B-lineage. This dataset includes 4 classes: Benign, Early Pre-B, Pre-B, and Pro-B Acute Lymphoblastic Leukemia.
Getting started is as simple as heading to and create your account or login. Once logged in you will be taken to the project selection/creation page.
Now it is time to import your data. You should have already downloaded the dataset from Kaggle, if not you can do so now on .
To deploy our model on our NVIDIA Jetson Nano, we are going to use the .
