Study finds generalist CT models struggle to tell liver tumors apart without extra training
A retrospective evaluation of two generalist CT foundation models found weak zero-shot differentiation of liver tumor types, with accuracy improving substantially after training only a linear classifier on frozen features.
In plain words
Generalist CT models are AI systems trained to detect many kinds of findings in medical scans. A new study tested how well two of them could identify which type of liver tumor a patient has. Without extra training the models did poorly, but when researchers added a simple classifier on top of the models' internal image features, accuracy rose sharply.
What was announced
The study, published as an arXiv abstract, evaluated RADAR and Merlin on a public single-institution dataset of 517 multiphase CT cases covering hepatocellular carcinoma, intrahepatic cholangiocarcinoma, breast and colorectal cancer liver metastases, and hemangioma. In zero-shot testing, meaning the models used their existing outputs with no task-specific training, RADAR reached 54.4% accuracy for four-class differentiation at the portal venous phase and misclassified 31% of malignant tumors as hemangioma, a benign lesion. Merlin reached 60.9% accuracy in that setting.
The authors then used linear probing, which trains only a linear classifier on the models' frozen image features while leaving the underlying model unchanged. With four-phase features, RADAR's accuracy rose to 79.9% and its misclassification of malignant tumors as hemangioma fell to 1%, while its agreement measure (kappa) improved from 0.37 to 0.72. The authors conclude that these generalist models have limited zero-shot ability to differentiate liver tumors, which improved substantially with linear probing alone. Evaluation used stratified fivefold cross-validation repeated five times.
What it does not show
- The evidence is an abstract only; full methods, additional results and peer review status are not available from this source.
- The dataset is a single-institution public cohort of 517 cases, so performance may not generalize to other populations or scanners.
- Only two models, RADAR and Merlin, were evaluated, and Merlin's linear-probing result is not reported in the available text.
Sources
Source-support review completed 20 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.
Coverage: arXiv: abstract.
Updates and corrections
No corrections so far.
First published 20 Nov 2026
New to the subject? AI from the Ground Up starts from the beginning.