As many of you might already know, malaria is a serious disease transmitted by the Anopheles mosquito. It is especially common in tropical regions and among people living near forests and hills. Alarmingly, malaria remains one of the leading causes of death in many parts of the world.
Alright! Let’s not waste any more time and dive right in.
According to the dataset we’ll be working with, there are a total of 27,558 cell images, neatly divided into two classes:
Infected – cells that show signs of malaria
Uninfected – healthy, non-infected cells
Before we jump into the modeling part, it’s important to understand what a CNN (Convolutional Neural Network) actually is.
CNN is a specialized type of neural network designed primarily for image data. Inspired by the way humans process visual information, CNNs are excellent at learning spatial hierarchies of features — meaning they start by learning simple shapes like edges and gradually combine them to understand more complex structures like textures or objects.
If you’re new to Neural Networks, I highly recommend reading our earlier post by P’Wynn explaining the basics in a very beginner-friendly way.
[Click here to read that post!]
A CNN typically consists of three main components:
Convolutional Layers – These use filters to detect patterns such as lines, corners, or textures in the image.
Pooling Layers – These reduce the size of the feature maps, helping the network generalize better and run faster.
Fully Connected Layers – These interpret the extracted features and make predictions, like classifying if a cell is infected or not.
Usually, convolutional and pooling layers are stacked alternately, gradually transforming the raw image into meaningful insights.
Imagine this: a convolution layer uses something called a filter (or kernel) to slide across an input image. This filter multiplies its values with the underlying pixel values to generate a new image (called a feature map). This resulting map highlights the brighter and darker regions — typically with values between 0 and 1, where 1 indicates the brightest area.
The filters act like little magnifying glasses that focus on specific visual patterns. By stacking several of these layers, CNNs learn what an infected cell "looks like" by themselves!
Before we train a CNN, we need to prepare the data properly. The dataset typically comes in two folders: Parasitized/ and Uninfected/. Each folder contains .png images of individual blood cells.
Here's how we can prepare the data:
🧪 Shuffle and Split the Dataset
By applying CNNs to malaria cell images, we’ve:
Preprocessed the image dataset
Built and trained a CNN for binary classification
Evaluated model performance
Visualized model focus using Grad-CAM
This is a great first step into medical AI and computer vision. With further tuning and more advanced architectures like ResNet or EfficientNet, you can push the accuracy even further!