Can an interpretable AI spot small liver cancers on CT? External testing is promising—but incomplete
A peer-reviewed multicentre study tested dLI-RADS on 133 external patients. Sensitivity reached 93%, versus 78% for radiologist-assigned LI-RADS, but only sensitivity could be compared externally and every lesion was manually segmented.
The short version
- What happened
- The retrospective study used 143 patients for training, 62 for internal evaluation and 133 for external evaluation, representing 451 lesions in total.
- Who it affects
- For people with small liver lesions, a reliable second reader could support earlier, more consistent review. False positives can also trigger anxiety, repeat imaging or invasive work-up, so sensitivity must be weighed against specificity and real clinical consequences.
- What remains unproven
- Retrospective study
Keep in mind
Retrospective study
Manual segmentation
Selected lesions under 40 mm
Predominantly male cohorts
Composite non-cancer reference
External LI-RADS comparison limited to sensitivity
No prospective workflow test
No patient outcomes
Early-access accepted article
Commercial research affiliations
Source analysis · 1 linked source. Primary analysed source date: 30 October 2026.
Read the full report · 3 min The evidence, methods, implications and limitations
The separate test hospital is the study’s strongest feature
Researchers combined neural-network predictions with handcrafted imaging measurements that mirror major LI-RADS features, then used logistic regression to estimate whether a manually outlined liver lesion was hepatocellular carcinoma. Training used 143 patients with 196 lesions; internal evaluation used 62 patients with 84 lesions; and a separate external cohort included 133 patients with 171 lesions. Lesions over 40 millimetres were excluded. Histology was the cancer reference, while non-cancer lesions used a composite reference standard.
In the internal set, dLI-RADS reached 80% lesion-level accuracy versus 74% for conventional LI-RADS and met the study’s five-percentage-point non-inferiority test. In the external set, sensitivity was 93% versus 78% for LI-RADS. The model’s external specificity was 74%, overall accuracy 86% and area under the curve 84%. A separate hospital test reduces the risk that the result depends only on one local scanner or workflow.[1]
It is not yet evidence that the system improves diagnosis or care
The external head-to-head comparison is narrower than the headline numbers suggest: LI-RADS features were unavailable for non-cancer lesions, so only sensitivity—not specificity or accuracy—could be compared directly. Every lesion was manually segmented, leaving an important labour and reproducibility step outside the model. The retrospective cohorts were also predominantly men, and the study did not test reporting time, reader behaviour, false-positive work-ups, treatment access or patient outcomes.
Several authors work at Guerbet Research, while the paper reports Région Île-de-France and ANRT funding and declares no competing interests. A prospective multi-reader study should now test automated or routine segmentation, prespecified thresholds, subgroup performance, calibration and downstream decisions. Until then, dLI-RADS is a promising externally tested decision-support method—not a replacement for radiologists or proof of better cancer outcomes.[1]
The distinction that matters
What the three cohorts establish
Separating development from evaluation helps show where the result is strongest and where evidence is still missing.
Training
143 patients; 196 lesions
Built the combined neural-network and imaging-feature model
Internal test
62 patients; 84 lesions
80% accuracy versus 74% for LI-RADS; non-inferiority met
External test
133 patients; 171 lesions
93% sensitivity, but only sensitivity was directly comparable with LI-RADS