New Tiny AI Spots Blood Cancer in Seconds — No Fancy Lab Needed

A researcher at North South University in Bangladesh has published a paper describing LukeNet, a small AI program that reads microscope photos of blood cells and identifies acute lymphoblastic leukemia (ALL) — a fast-growing blood cancer that is the most common childhood cancer. The idea is simple: the earlier doctors catch it, the better the odds. A computer can scan images far faster than a human eye and never gets tired. The paper is a preprint, meaning it has not yet been checked by other scientists.

What makes LukeNet unusual is its size. Most AI systems that read medical images are huge and need powerful computers. LukeNet was designed to run on small, low-cost gadgets — the kind of connected health devices, or wearables, that track temperature, heart rate and oxygen. That matters because many clinics, especially in poorer regions, don't have fancy computers or specialist pathologists on hand. If the AI runs on a phone or a bedside device, a blood scan could be checked in minutes rather than sent away for days.

The reported numbers are striking. Trained on three separate leukemia image datasets, LukeNet claimed 99% accuracy on data it had seen and 99% on images it had never seen before. That beats six famous AI models, including ResNet50, VGG16 and MobileNet, plus two combined 'ensemble' models. The paper also adds explainable AI (AI that shows its reasoning) so it highlights the exact regions of the image it based its decision on. That helps doctors trust and double-check the result instead of accepting a black-box answer.

The catch: this is a lab result, not a hospital result. Test datasets are cleaned-up and carefully labelled; real patient samples are messier. The system is also not approved by any medical regulator and hasn't been tested on actual patients in a clinic. Treat the 99% as promising, not proven.