R-Super Framework Uses Radiology Reports to Train Tumor Segmentation AI, Boosting Multi-Cancer Detection

AI trained on 127,496 CT reports (R-Super) detected 56% more malignant tumors than radiologists and improved sensitivity by over 11% beyond mask-only training, enabling multi-cancer detection without scarce tumor masks.

• Training on >100,000 reports without any mask outperformed mask-only models trained on 870 masks.
• Combining reports and masks increased cancer detection sensitivity by +11% and Dice score by +14% beyond mask-only training.
• External validation across four sites (USA, Turkey, Switzerland) covered 7 tumor types: spleen, gallbladder, prostate, bladder, uterus, esophagus, and adrenal.

Automated summary

RadPigeon summaries are original and for information only. They are not clinical advice.