Detecting sickle cells using AI

The goal

We were set out to create an AI that uses object detection to detect sicklecells in microscope images in the different stages of deoxygenating. The primary goal was to create a highly accurate model, we were aiming of an accuracy of around 90%.

Annotating the dataset

The first step we took was to annotate the images that the client had provided us. We made a selection from the images that we had. To ensure that there was no unintentional bias on our end these where random picked using code.

After selecting the images, we annotated the images in Label stadio, which allows us to manually draw bounding boxes around the sickle cells. These annotations can then be used to train and validate the model. In total we annotated around 570 images, each containing several sickle cells.

Shows how to annotate images in Label studio

Converting our annototations to Detectron2

The annotations exported by label studio were not in the correct format for Detectron2 yet, to ensure that the annotations were done correctly we used a python script to load the JSON and convert the data to the correct format for Detectron2.

Code to format the mini JSON to Detectron2 format.

Optimizing the modal for the highest accuracy

As we wished to gain a very high accuracy, we went through several different itterations of the modal. For this we greatly tweaked the settings of the modal. Until we reached the highest accuracy without the modal becoming overfit.

Showing the annotations the modal made

Since we were creating a medical AI, a high accuracy was the most important part of the AI. Since the AI was primarily aimed at being used for research purposes that can be build upon in future projects. The majority of our time was spend annotating images and optimizing the modal as much as possible.

Beside the accuracy score we also hand validated the annotations by writing code that shows us the annotations the modal has made. Which allowed us to further tweak the modal according to our findings.

Code to show the output of the modal to the user, this is code that runs on validation images.

An annotation of the sicklecells made by the modal, the modal has annotated 5 sicklecells, of which 4 are 99% certaincy and 1 is 100% certaincy.

The end result

The final produced modal for Dectecton 2 was trained with 420 training images, 50 validation images and 50 test images. In the image to the right you can see the confusion matrix produced for the Detectron2 modal.

The Detectron2 modal had the following results:

Accurcy 95.3%
F1-Score 85.1%
PR-Score 92.3%

Confusion matrix for the detectron2 modal, showing the modal correctly predicted 262 sicklecells, wrongly marked 22 background objects as sickle cells and wrong marked 14 sickle cells as background.