Summary and Setup

Introduction

This lesson introduces the core concepts of artificial neural networks through a practical application in healthcare: training a model to classify chest X-ray images.

You’ll learn how to:

  • Load and visualize medical imaging data
  • Prepare images for use in machine learning
  • Build and train a convolutional neural network
  • Evaluate model performance
  • Explore explainability techniques using saliency maps

By the end of the lesson, you’ll have constructed a neural network capable of detecting pleural effusion in chest X-rays — a real-world example of how machine learning can assist clinical decision-making.

Prerequisite

Prerequisites

You need to understand the basics of Python before tackling this lesson. The lesson sometimes references Jupyter Notebook although you can use any Python interpreter mentioned below in the setup instructions.

Getting Started

To get started, follow the setup instructions mentioned to download data and install a Python interpreter.

Software Setup

All of the software and data used in this lesson are freely available online, and instructions on how to obtain them are provided below.

Install Python

In this lesson, we will be using Python 3 with some of its most popular scientific libraries. The simplest way to install Python is probably to install the uv package manager following the instructions from ORNL-training’s lesson on Python with uv.

Open a Terminal and Create a Project Directory

We will keep track of all our source code and data files in a single “project” directory, Desktop/src/ml-xray. In order to run shell commands from that directory we first have to open a shell and navigate there.

If you’re using a Unix shell application, such as Terminal app in macOS, Console or Terminal in Linux, or Git Bash on Windows, execute the following command:

BASH 

cd ~/Desktop
mkdir src
cd src
uv init ml-xray
cd ~/Desktop/src/ml-xray

On Windows, you can use its native Command Prompt program. The easiest way to start it up is pressing Windows Logo Key+R, entering cmd, and hitting Return. In the Command Prompt, use the following commands to create your project:

cd /D %userprofile%\Desktop
mkdir src
cd src
uv init ml-xray
cd /D %userprofile%\Desktop\src\ml-xray

Download Project Package Dependencies

The major strength of Python is the number of libraries available. These provide functionality like loading and manipulating images, organizing and visualizing tables of data, building and training AI/ML models etc. This lesson is built on several Python libraries.

We can install them up-front using the following commands:

BASH 

uv add opencv-python
uv add matplotlib
uv add numpy
uv add keras
uv add notebook # for using jupyter (recommended)

The Python Package Index hosts most python packages. Any developer can add packages there, so it is very important to read through and understand the package’s origin, contribution history, and developers.

Packages do not always have the same name as the corresponding import source code line (e.g. import cv2 for opencv-python). Digging through available packages like this requires effort, but not as much effort as writing the package yourself. Often, the package’s documentation is the deciding factor for how successful and useful it will be.

When you want to progress beyond this lesson, the package documentation and the wider set of materials written about using these packages are the place to go.

Obtain lesson materials

The data that we are going to use for this project consists of 700 chest X-rays. These X-rays are a subset of the public NIH ChestX-ray dataset.

Xiaosong Wang, Yifan Peng, Le Lu, Zhiyong Lu, Mohammadhadi Bagheri, Ronald Summers, ChestX-ray8: Hospital-scale Chest X-ray Database and Benchmarks on Weakly-Supervised Classification and Localization of Common Thorax Diseases, IEEE CVPR, pp. 3462-3471, 2017

  1. Download chest_xrays.zip.
  2. Unzip the downloaded files to Desktop/src/ml-xray.

After unzipping, a new folder (chest-xrays) should be present in the project directory. It has two sub-folders, each full of png images.

Opening a Jupyter Notebook

A Jupyter Notebook provides a browser-based interface for working with Python. We will use notebooks in this tutorial, but it is possible for an expert to avoid notebooks by typing code directly into any editor (e.g. vim main.py) and running (uv run main.py).

The advantage of editing main.py is that you will tend to structure your code into functions. Ideally, functions accomplish small units of work and have only a few variables defined. Jupyter notebooks are good for exploring ideas, but easily become a confusing jumble of variables (holding lots of data). Best practice is actually to mix notebooks with python modules (e.g. helpers.py, ml-model.py, etc.). That way, you can write re-usable functions in your py files and import helpers from your Jupyter notebook.

BASH 

uv run jupyter notebook
uv run python -m notebook

Launch the notebook by clicking on the “New” button on the right and selecting “Python 3” from the drop-down menu: Anaconda Navigator Notebook directory