Diffusion model
A generative model that creates images, and increasingly video and 3D volumes, by starting from random noise and removing it step by step until a realistic picture remains. It learns this by being shown real images with noise added and trained to undo it, so what it generates follows the patterns of its training images, and it can sometimes reproduce one of them closely. In medicine it is used to make synthetic training images, to denoise scans and to help reconstruct images from less data. A synthetic image shows plausible anatomy that belongs to no actual patient unless it closely copies a training image, and on a real patient's scan the model can add realistic detail that is not there or smooth away a real finding.
A vendor explains that its fracture detection model saw too few pediatric elbow injuries, so it added thousands of synthetic radiographs made with a diffusion model to the training set. That can help a model learn rare patterns, and the generated images can also carry anatomy that looks right and is subtly wrong, such as ossification centers that do not match the child's age. Ask what share of the training data was synthetic, and confirm that every performance figure you are shown comes from real patient images kept apart from training.
Terms like this come up in real clinical scenarios across the HelloAI courses: bite-sized modules with verifiable certificates. An account takes one minute, no password needed.
Sign in →