Medical AI can now learn from unlabeled CT scans by understanding their 3D structure instead of treating them as flat pictures

Researchers built a new type of self-supervised learning system that processes 3D medical scans (like CT images) as actual volumes instead of slicing them into 2D pictures and processing each slice separately. This means the system learns spatial relationships that exist in 3D space, which is closer to how human radiologists actually understand medical images, and it performs better on reconstruction tasks without requiring labeled training data.

Medical AI systems need labeled data to learn, but labeling CT scans is expensive and slow. Most existing systems get around this by pre-training on natural images (like photos of cats) or by treating 3D scans as stacks of flat pictures, both of which throw away critical information about how structures connect in three-dimensional space. This approach captures that 3D context without requiring labels, which could make it faster and cheaper to build medical AI systems that actually understand what they're looking at.