Detecting Pneumonia with Deep Learning: A Soft Introduction to Convolutional Neural Networks

•4 min read

With the rise in popularity of neural networks, researchers and engineers have found world-changing applications for computer vision. Deep learning now allows us to automate analysis techniques once thought impossible for computers.

Today, I want to show you an example. By acquiring a labeled dataset of chest x-ray scans for pneumonia, I was able to rapidly build and prototype an artificial intelligence model that identified pneumonia cases with 80% accuracy, all within an afternoon.

Pneumonia is a common disease we’ve fought for thousands of years. While preventative care has improved in the West, many cases still occur throughout Africa and Asia Minor. Automating detection would greatly improve the efficiency of radiologists, letting them focus on edge cases needing expert judgment.

So… Where to Start?

Computer Vision & Convolutional Neural Networks

If we want to automate the detection of pneumonia from x-ray scans, we’re dealing with a computer vision problem. Our goal is to teach software to interpret and analyze image data.

To do this efficiently, we use Convolutional Neural Networks (CNNs)—the current standard in computer vision. CNNs are similar to traditional feed-forward neural networks, but they implement a special mathematical operation called a convolution.

So what does that mean in plain English?

CNNs are far more efficient at processing images than traditional neural networks.

Instead of examining each pixel independently, CNNs apply filters to regions of the image. This greatly reduces the number of operations—down from hundreds of millions to under ten million in some cases.

Organizing Our X-Ray Scans

Thanks to Kaggle, I obtained a labeled dataset of chest x-rays, divided into two folders: one for "Normal Scans" and one for "Pneumonia Scans".

Detecting pneumonia here is a binary classification problem: either we detect it (1), or we don’t (0). I used Python to label and shuffle the images into a unified dataset.

While the dataset was relatively clean, it was still imbalanced—about 3x more pneumonia images than normal. I also had to resize all images to a uniform 224x224 RGB format.

Working with real-world data is always a challenge—but that’s a conversation for another time.

Building Our Model

Deep learning works by analyzing data through multiple layers. The more layers, the more complex the tasks we can tackle—hence, "deep" learning.

To get started quickly, I used Keras, a high-level neural network API, with TensorFlow as the backend.

Here’s what a basic CNN model looks like using Keras:

from tensorflow.keras.models import Sequential
from tensorflow.keras.layers import Conv2D, MaxPooling2D, Flatten, Dense, Dropout

model = Sequential([
    Conv2D(32, (3, 3), activation='relu', input_shape=(224, 224, 3)),
    MaxPooling2D(pool_size=(2, 2)),
    
    Conv2D(64, (3, 3), activation='relu'),
    MaxPooling2D(pool_size=(2, 2)),
    
    Conv2D(128, (3, 3), activation='relu'),
    MaxPooling2D(pool_size=(2, 2)),
    
    Flatten(),
    Dense(128, activation='relu'),
    Dropout(0.5),
    Dense(1, activation='sigmoid')  # Binary classification: Pneumonia (1) or Normal (0)
])

model.compile(optimizer='adam',
              loss='binary_crossentropy',
              metrics=['accuracy'])

Even without deep technical knowledge, you can see each line adds a layer to the network. This simplicity is why Keras is so great for rapid prototyping.

Training and Testing the Network Now for the fun part—training the model!

Here’s what happens behind the scenes:

The network starts with random weights and zero bias.

It processes an image through all the layers and makes a prediction (0 or 1).

It compares the prediction with the true label using a loss function.

It uses backpropagation (just fancy calculus) to adjust weights and biases to improve predictions.

We repeat this process across all training images, aiming to reduce the loss over time.

Here’s what sample output might look like during training and testing:

Epoch 1/10
100/100 [==============================] - 25s 240ms/step - loss: 0.6524 - accuracy: 0.6200
Epoch 2/10
100/100 [==============================] - 24s 230ms/step - loss: 0.4381 - accuracy: 0.7850
Epoch 3/10
100/100 [==============================] - 24s 235ms/step - loss: 0.3652 - accuracy: 0.8300
...
Epoch 10/10
100/100 [==============================] - 23s 225ms/step - loss: 0.2104 - accuracy: 0.9150

Evaluating on test set...
25/25 [==============================] - 2s 65ms/step - loss: 0.3701 - accuracy: 0.8000

Test Accuracy: 80.00%

80% accuracy! Not bad for a quick first run.

Some Concluding Thoughts While this pneumonia detector isn’t a unicorn product that will change the world tomorrow, it shows just how accessible and powerful deep learning has become.

Setting up a CNN was quick and painless. Training on a decent GPU took less than an hour.

That’s the beauty of deep learning today—the barriers to entry are lower than ever.

Whether you're technical or just curious, I hope this gave you insight into how modern computer vision works. And if you're interested in getting started, don’t hesitate to reach out—I'd be happy to help you take the first step.