AI diagnosis from medical images: Science
Principles
The science it rests on, with why each principle matters to it
Relies on- Artificial neural networks The software is built from layers of simple artificial neurons that turn pixel values into a diagnosis [1].
- Backpropagation and gradient learning Networks learn from graded images by backpropagation, which adjusts every weight to reduce the error on each example.
- Probability and statistics Outputs are scores, and performance is judged by how often disease is correctly found and correctly ruled out [3].
as of 2026-11-11
Discoveries
The discoveries that made it possible: when, by whom and what was found
| Year | Who | What |
|---|---|---|
| 2012 | Dan Ciresan, Alessandro Giusti, Luca Gambardella and Jürgen Schmidhuber | A deep network finds dividing cells in breast cancer tissue images better than other methods [1] |
| 2016 | Varun Gulshan, Lily Peng and colleagues at Google | A deep network grades diabetic eye disease in retinal photographs about as well as ophthalmologists [2, 7] |
| 2023 | Kristina Lång and colleagues | A randomised trial's early analysis shows AI-supported breast screening finds a similar rate of cancers with less reading work [5] |
as of 2026-11-11
The science explained
The science behind it explained for a curious reader
Imagine teaching a trainee by showing a stack of thousands of scans, each with an expert's verdict written on the back. The trainee guesses, turns the card over and adjusts. A neural network does the same with arithmetic. Each image enters as a grid of numbers. Layers of artificial neurons combine neighbouring numbers, so the first layers pick out edges and spots and later ones assemble them into features such as a small bleed in the retina. Every wrong guess sends a correction back through the layers, the method called backpropagation. After many thousands of examples the network has tuned a great many weights. It has no idea what an eye is, but it has learned which patterns of pixels go with which verdicts, which is why it must be tested on new images before use.
as of 2026-11-11
Sources
- Mitosis Detection in Breast Cancer Histology Images with Deep Neural Networks MICCAI 2013, Lecture Notes in Computer Science 8150, pp. 411 to 418, Springer (D. C. Ciresan, A. Giusti, L. M. Gambardella and J. Schmidhuber, IDSIA, Lugano), 2013
- Deep Learning for Detection of Diabetic Eye Disease (29 November 2016) Google Research blog (L. Peng and V. Gulshan), 2016
- FDA permits marketing of artificial intelligence-based device to detect certain diabetes-related eye problems Healthcare Innovation (HCI Innovation Group), reproducing the FDA news release of 11 April 2018, 2018
- Artificial intelligence-supported screen reading versus standard double reading in the Mammography Screening with Artificial Intelligence trial (MASAI) The Lancet Oncology 24(8), 936 to 944 (K. Lång and others), record at Lund University, 2023
- Development and Validation of a Deep Learning Algorithm for Detection of Diabetic Retinopathy in Retinal Fundus Photographs JAMA 316(22), 2402 to 2410 (V. Gulshan and others), record in the BVS/LILACS database, 2016