AI scans and liver cancer: a survival signal, not a treatment rule
AI measurements from routine scans were associated with liver-cancer survival. The retrospective study does not establish which treatment a patient should receive.
The short version
- What happened
- The retrospective study analysed 1,512 patients with hepatocellular carcinoma across 10 cohorts at six Chinese centres; 986 died and 1,298 progressed or died during follow-up.
- Who it affects
- Patients could eventually gain more prognostic information from scans they already undergo, but premature use could wrongly change a cancer treatment. The study measured associations, not better treatment choices or quality of life.
- What remains unproven
- Retrospective Chinese cohorts; 84.1% male and 82.5% hepatitis-B positive; residual confounding; inconsistent treatment-line data; locally assessed progression; unpaired CT and MRI; no prospective test of added clinical value.
Keep in mind
Retrospective Chinese cohorts; 84.1% male and 82.5% hepatitis-B positive; residual confounding; inconsistent treatment-line data; locally assessed progression; unpaired CT and MRI; no prospective test of added clinical value.
Source analysis · 1 linked source. Primary analysed source date: 17 October 2026.
Read the full report · 3 min The evidence, methods, implications and limitations
AI extracted more information from scans already taken
Researchers used a pretrained deep-learning system to segment nine abdominal organs or structures on pretreatment CT or MRI scans. The 1,512 adults came from 10 cohorts at six Chinese centres; 84.1% were men, 80.8% had advanced BCLC stage C disease and 82.5% had hepatitis B. Median follow-up was 32.03 months.
The team tested 18 absolute and height-normalised measurements. Models adjusted for age, sex, body mass index, cirrhosis, hepatitis B, creatinine, tumour burden, metastasis, treatment regimen and scan type, then corrected for multiple comparisons. A 100-scan subset was checked against expert-corrected organ masks.[1]
Why the result is not a clinical decision rule
Three continuous measures remained associated with overall survival: larger liver volume and left-adrenal volume tracked lower mortality, while a larger height-normalised right kidney tracked higher mortality. No continuous measure survived correction for progression-free survival. In a separate threshold analysis, no organ measure retained a corrected association with overall survival.
This was a retrospective observational study. Treatment line was not consistently recorded, CT and MRI came from different people, and progression was assessed locally on varying schedules. Organ size may reflect disease, physiology or treatment selection rather than cause an outcome, and the study did not show that the measures improve decisions beyond established scores.[1]
The distinction that matters
From routine scan to a cautious survival association
The study reused pretreatment images, then tested whether automated organ measurements tracked later outcomes.
Measure
Segment nine organs on CT or MRI
Create 18 absolute and height-normalised features
Compare
Adjust survival models for clinical and tumour factors
Three features retain an overall-survival association
Do not infer
Avoid treating association as treatment response
Prospective decision value remains unproven
What this could mean for patients and imaging teams
If independently validated, automated organ measurements could add a low-burden layer to scans patients already receive, without another examination or radiation dose. For now, clinicians should not use these cut-offs to start, stop or choose immunotherapy.
The assessment would change with prospective external cohorts, paired evaluation across scan types, consistent treatment-line data and proof that the measurements improve calibrated predictions and patient decisions beyond tumour stage, liver function and routine laboratory results.[1]