Exploring AI in healthcare: Kian van Holst's image classification project

Can you describe the project that you have been working on? 

Kian: ‘For my assignment, I was tasked with creating an image classification system. I had the option to choose between binary (two options) or multiclass classification (three or more options). I decided to challenge myself and went with a system that could handle three classes.’ 

‘I have an interest in exploring how AI can contribute to healthcare. Therefore, I chose to develop an image classification system specifically for healthcare applications. The system would classify images of healthy lungs, lungs with bacterial infections, and lungs with viral infections. To gather the necessary data, I accessed a dataset from Kaggle, a data science platform and community.’ 
 

What did the process of creating the image classification system look like?  

Kian: ‘I started by establishing baselines for my project. What is the level of human performance, what is the random guess? This can help to determine whether a model can be benefitted from. Then, I focused on data pre-processing, ensuring that the system could process the data for training. To analyse the images, I employed Convolutional Neural Networks (CNN), which are neural networks (NN) capable of processing 2D images. By creating layers that act as filters, I aimed to identify and highlight key differences in the images, connecting them to specific outcomes.’  

‘We primarily worked with Python and utilized various packages and modules for importing and processing the data. This method is an example of a supervised learning algorithm. This involved providing the computer with labelled examples of each image category—longs that are healthy, longs with a bacterial infection, and longs with a viral infection. The system learned from these examples to classify new, unseen images. After completing the model, I assessed its performance, achieving an accuracy of 83% in my three-class classification.’ 

‘To improve my project, I employed a transfer learning method, which involves leveraging pre-trained models and integrating them with our own model. This approach significantly reduced the training time required for the system.’ 
 

Did you encounter any challenges while working on this project? 

Kian: 'During the analysis of the confusion matrix, I noticed that differentiating between bacterial and viral infections proved to be more difficult compared to distinguishing healthy and unhealthy lungs. The system achieved an accuracy of 94% in that binary (healthy / unhealthy) classification. Since identifying whether someone needs assistance is of utmost importance, the distinction between bacterial and viral infections, although challenging, became less critical.’ 

‘AI's potential extends across various fields, but luckily with the ADS&AI programme at BUas we are free to chose our topics and pursue our personal interests. This allows us as students to gain insights that will guide us in our future careers.’ 

How can this model be implemented practically in the field of radiology? 

Kian: ‘Aside from use in radiology, one potential application for the model is an app designed to assist radiology students in their studies (like the prototype I made). By implementing the image classification system and explainable AI methods, like GradCAMs (which are heatmaps over the image, to show what the model looks at), students can learn to identify different lung conditions accurately. This project not only allowed me to explore the practical implementation of AI models but also helped me develop a deeper business understanding how AI can be implemented in several ways.’ 

 

If you would like to know more about Kian, visit his LinkedIn profile.