Related Experiment Videos
Large-Scale Multi-Cancer Detection by Learning Segmentation from Reports
Pedro Bassi1,2,3, Xinze Zhou1, Wenxuan Li1,2,3
1Johns Hopkins University, Baltimore, MD, USA.
Research Square
|July 29, 2026
Summary
This study introduces R-Super, an AI framework that uses radiology reports to train tumor segmentation models, overcoming the scarcity of manual tumor masks. R-Super significantly improves cancer detection accuracy, even surpassing radiologists in some cases.
Area of Science:
- Artificial Intelligence in Medical Imaging
- Radiology and Pathology Report Analysis
- Computational Pathology
Background:
- Computed tomography (CT) scans are widely used, but early tumor detection is challenging due to limitations in current AI models.
- Segmentation-based AI models require tumor masks, which are scarce and costly, hindering widespread application.
- Radiology reports contain valuable tumor descriptions but are underutilized for training AI segmentation models.
Purpose of the Study:
- To develop a novel framework, R-Super, that leverages radiology and pathology reports to train AI for tumor segmentation.
- To overcome the limitations of scarce tumor masks by utilizing readily available report data.
- To improve the accuracy and scalability of AI-driven cancer detection in CT scans.
Main Methods:
- R-Super framework converts routine radiology and pathology reports into localized training signals for AI segmentation models.
- Trained on a large dataset of 127,496 CT-Report pairs (42 million images).
- Evaluated performance across multiple institutions in the USA, Turkey, and Switzerland, comparing against mask-based models, alternative frameworks, and radiologists.
Main Results:
- R-Super successfully detects 7 tumor types with scarce public mask data (spleen, gallbladder, prostate, bladder, uterus, esophagus, adrenal).
- Report-based training with R-Super outperformed mask-based models trained on significantly more masks.
- Combined report and mask training with R-Super increased cancer detection sensitivity by +11% and DSC by +14% compared to mask-only training.
- R-Super surpassed 6 alternative training frameworks, 9 leading public AI models, and 6 out of 7 radiologists in detection accuracy, identifying 56% more malignant tumors on average.
Conclusions:
- Radiology reports are a valuable, underutilized resource for training accurate cancer detection AI.
- R-Super enables AI segmentation models to scale beyond limited manual annotations, advancing automated cancer detection.
- The R-Super framework and associated code/data release facilitate clinical deployment of advanced AI for incidental and automated cancer detection.