Johns Hopkins releases Merlin-Cancer-Super, a CT tumor segmentation model trained from radiology reports for 9 organs

Researchers at Johns Hopkins University have released Merlin-Cancer-Super on Hugging Face, an AI model that detects and segments tumors on CT scans in nine organs: the bladder, gallbladder, spleen, esophagus, stomach, duodenum, prostate, uterus and adrenal glands. The model comes out of the R-Super project at Johns Hopkins, whose papers list Zongwei Zhou as corresponding author, in collaboration with the Merlin project at Stanford. The problem it addresses is the cost of training data. Tumor segmentation models usually need per-voxel tumor masks drawn by experts, which are scarce, while hospitals hold large archives of radiology reports that already describe the tumors found on each scan. Report Supervision (R-Super) is a training method that turns those reports into direct supervision. Before training, a large language model extracts tumor information from each radiology report. R-Super then adds two new loss functions, a Volume Loss and a Ball Loss, which use the extracted information to teach the segmentation model to produce tumors that agree with the report in tumor count, diameters and locations. The method was presented at MICCAI 2025, where its paper, "Learning Segmentation from Radiology Reports", was runner-up for the Best Paper Award (top 2 of 1,027 papers). Merlin-Cancer-Super is R-Super trained on the radiology reports and tumor masks of Merlin Plus, a dataset accepted at MICCAI 2026. Merlin Plus adds longitudinal metadata (patient IDs and scan dates) and per-voxel annotations of organs and nine tumor types to Stanford's Merlin dataset of 25,494 CT scans; it is published on Hugging Face. According to the Merlin Plus paper, the model surpasses state-of-the-art public AI models and leading vision-language models in tumor detection and segmentation in all nine organs. The released model takes CT volumes in NIfTI format with a list of case IDs and outputs, for each scan, segmentation masks for the nine tumor types and 26 organs, with optional tumor probability maps; inference can be split across several GPUs. Code is available on GitHub, the model is licensed CC BY-NC-ND 4.0, and the work was supported by the Lustgarten Foundation for Pancreatic Cancer Research, the Patrick J. McGovern Foundation and the US National Institutes of Health.