Deep learning uses neural networks with many stacked layers. Each layer transforms its input, and the network learns useful representations at increasing levels of abstraction without a human defining what to measure.

This is what changed medical imaging AI. Earlier computer-aided detection required engineers to hand-specify features like edge sharpness or texture measures. Deep learning removed that step, which is why performance improved sharply after roughly 2012 and why nearly every device on the FDA list authorized since 2017 uses it.

Where This Gets Misread

Learning its own features is also why the model cannot tell you what it learned. Hand-engineered features were auditable by definition. Learned representations are not, and the resulting opacity is a direct consequence of the thing that made the method work.