Computer vision: how does a machine spot a tumour in an image?

To a radiologist, this is an image. To a machine, it's just a grid of numbers. So how does it spot a tumour? Around 1959, researchers found neurons in a cat's visual cortex that respond to lines at a precise orientation. In 1989, AT&T researchers applied a convolutional network to reading handwritten zip codes supplied by the US Postal Service. In 2012, a deep convolutional network won the ImageNet competition, with a clear lead over its rivals. It's called computer vision. Three steps. Let's go. Step one: the image becomes numbers. In this photo, every pixel is a value: zero for black, two hundred and fifty-five for white. A scan holds hundreds of thousands of them. Before any maths, they're normalised: same scale, same contrast. Step two: the network looks for patterns. A small filter slides across the image and lights up where it finds a shape: here, the edges. Layer after layer, this convolutional network turns edges into shapes, then objects. Vision Transformers take in the whole image at once. Step three: the model suggests, the radiologist decides. The system circles suspicious areas, each with a confidence score. The radiologist confirms the real alerts and dismisses the false ones, which can then help train the model. Google Health, Lunit, Aidoc and Siemens: on specific tasks, these tools already help spot breast and lung cancers. KOAEE assistants don't read images. But they follow the same rule: AI helps, people decide. A medication question at a clinic? The assistant refers you to a healthcare professional. And every screen says it: AI can make mistakes. To understand AI in 150 seconds, subscribe to the channel. The AI that stays between you. Try it at koaee.ai.