About the competition

Aneurysms are a bulge or a small balloon in the wall of a blood vessel. These can rupture randomly, causing bleeding inside the body. If this happens in the brain, death is a common outcome. Treatment for aneurysms already exists; however, aneurysms are generally symptomless, and the detection for half of the cases is only after they have ruptured. These aneurysms range from big (>1 cm) to very small (<3 mm) and can most commonly be found by analysis of an MRI t1 and t2 scan, but also by CTA or MRA scans. The goal of this competition was to detect if a patient has an aneurysm and identify where in the brain it is located based on scans of these 4 modalities. 

Radiological Society of North America

The Radiological Society of North America (RSNA) is a nonprofit organization focused on using radiology and AI for various medical applications. RSNA hosts machine learning competitions on Kaggle frequently, usually about using radiology scans (MRI or CT) to detect various diseases, like cancers or pneumonia. 

Relevance

3% of the entire world population develops an aneurysm, and over half are detected after they rupture. 500,000 deaths per year are caused by aneurysms in the brain, with half of the victims being younger than 50. These aneurysms are typically asymptomatic until they rupture. Typically a CTA scan is used to find an aneurysm; however, not all patients have a CTA scan available. MRI t1/t2 scans are a lot more commonly used for other purposes, so the problem this competition aims to solve is to develop a model that can take any modality of brain scan and locate an aneurysm.

Technical details 

Aside from the brain scans, the data also included various metadata, like patient and scanner information, and masks for the 13 brain vessels. Location information only consisted of the center of the aneurysm, not any information about the size or shape. Challenges included that some scans consisted of only the brain portion, while other scans started at the lungs and ended at the top of the skull. The aim of the model is thus to work with a lot of changing data, with different modality scans, different rotations, and different shapes of scans. 

Our final submission consisted of normalizing the data to be 1 mm³, creating a sphere with a radius of 5 mm around the center of the aneurysm, and using that sphere to train a segmentation model with 2 heads to simultaneously learn the location of the aneurysm in the scan while also learning to which binary label that belonged.

UN Sustainable Development Goals

This competition contributes to the UN sustainable development goal 3, good health and well-being, by attempting to find aneurysms earlier and allowing for better treatment for hundreds of millions of people.

Competition Page

Read More about this competition