Detecting Brain Tumors with AI

Third-year Applied Data Science & AI students Francisco Ribeiro Mansilha and Lea Banovac developed BrainScan AI, a proof-of-concept web application showing how artificial intelligence can assist in detecting brain tumors from MRI scans. The tool gives medical professionals and researchers an interface to analyse brain images and generate AI-powered predictions across three tumor types.

Where the idea came from

Both students share an interest in computer vision and healthcare. Francisco found a professionally labelled MRI dataset and recognised the potential for a meaningful medical application. Lea saw it as a chance to combine those interests while building a strong portfolio piece ahead of internship applications.

How they built it

The work ran in three phases: analysing the data and selecting a model, training and refining it with explainable AI visualisations, and turning the trained model into a web application usable by non-programmers.

The team used a YOLO (You Only Look Once) algorithm fine-tuned for tumor detection. Their dataset held 5,249 images labelled across four categories — Glioma, Meningioma, No Tumor and Pituitary. The model analyses sections of an image, identifies tumors, and returns a classification, location data and a confidence score. Training used lateral, superior and anterior MRI views so the model would hold up across perspectives.

Explainability

Lea prioritised explainable AI, highlighting the regions of an image that drove each prediction. As Francisco puts it:

“Without explainable AI, you can’t build trust with medical professionals.”

That transparency matters most in exactly this kind of setting, where understanding how a decision was reached is part of the decision.

The application

The app has three main sections: an overview page with links to the dataset, an image analysis view with contrast, brightness and rotation controls, and a detection view where a scan can be uploaded to generate predictions with confidence scores and a downloadable overlay image.

Limitations and what’s next

The main constraint is data. Francisco notes that real-world deployment would need “a much more diverse set of images and labels” — the system currently recognises three tumor classes or their absence. The pair see BrainScan AI as a foundation that could grow through collaboration with medical institutions, both to improve accuracy and to cover more tumor types.

Their second-year computer vision coursework and explainable AI training laid the groundwork; earlier projects such as plant root segmentation built the technical skills this drew on.