New AI Learns from Doctors' Reports to Find Tumors Better

A new AI system called Report Supervision, or R-Super, could make tumor detection more accurate by learning from something doctors already produce: written radiology reports. Typically, AI tumor-detection models are trained on "masks" — detailed outlines of tumors that specialists draw by hand. Creating just one 3D mask can take 30 minutes, which is why even the largest datasets only have a few thousand. That's a bottleneck.

R-Super solves this by using the text reports that radiologists write every day. These reports describe a tumor's size, location, and how many there are. The researchers taught the AI to compare its own findings with the written descriptions during training. This way, the AI learns to match what a professional would actually note, without needing thousands of hand-drawn outlines.

The results were significant. On external tests, R-Super improved detection accuracy and segmentation quality by up to 15% compared to training on masks alone. It also beat other methods like CLIP and multitask learning. Most importantly, it helped most when training data was scarce — for example, with just 50 masks — while also improving performance with thousands of masks. This means hospitals with limited resources could potentially build much stronger AI tools.

The catch: this is a research breakthrough, not a product yet. The reports are used only during training, not when the AI is used in practice. And the system has been tested on kidney and pancreatic tumors, not all cancers. Still, using existing medical notes to train better AI is a promising, low-cost way to bring machine learning into everyday diagnosis.