Multi-lesion radiomics predicts 12-month chemo response in liver metastases: AUC 0.79, but with 32 patients

By Bioengineer.org · October 18, 2026.

Researchers in Guangzhou (China) have developed a machine learning model that estimates, before the first infusion, whether a patient with unresectable colorectal liver metastases will remain progression-free 12 months after starting first-line chemotherapy. According to the study, published in BMC Medical Imaging, the model achieves an AUC of 0.792 on the internal validation set.

The AUC (area under the ROC curve) measures the ability to distinguish patients who progress from those who don't: 0.5 is chance and 1.0 is perfect. Here the 95% confidence interval runs from 0.625 to 0.933, a range spanning from modest to excellent performance.

The foundation is radiomics: converting a tumor's pixels into hundreds of quantitative variables (texture, gray-level distribution, relationships between pixels) that reflect biology invisible to the eye. The novelty is that the team analyzes all measurable metastases in each patient, not just the largest, as most previous studies did. This captures the heterogeneity between lesions, which may have prognostic value.

The retrospective cohort includes 159 patients, 127 for training and 32 for internal validation. Radiologists segmented the lesions in two phases of contrast-enhanced CT: arterial and portal venous. The final model, called AVC (arterial, venous and clinical), combines both phases with the tumor markers CA19-9 and CEA and with whether the primary tumor had been resected. It integrates several classifiers through voting weighted by their training performance.

To select variables without memorizing noise, a common problem when there are more variables than patients, the authors use a method called CRITIC (weighting criteria by their contrast and redundancy). They integrate it into a multi-criteria decision-making framework. It achieved the highest numerical AUC compared with other selection techniques, although the differences were not statistically significant. For interpretability they use SHAP, which quantifies how much each variable contributes to each prediction. Intensity and texture features from both phases carry weight, as do CA19-9, CEA and resection of the primary tumor.

The authors themselves acknowledge the limitations. The study comes from a single center, is retrospective and is underpowered to compare models against each other. Decision curve analysis suggests net benefit across part of the relevant threshold range, but requires confirmation in larger external cohorts.

**Our take.** What's valuable here is not the 0.79, which with 32 patients is almost anecdotal, but the design. It starts from a CT scan and lab tests already performed in any hospital, so it requires no new scanners or biopsies. It is also explainable and respects the real heterogeneity of the disease. That combination of low deployment cost and interpretability is what tends to survive the transition to the clinic.

The risk is the usual one in radiomics: models that look promising at one center lose performance when the scanner, protocol or population changes. In general, most of these models fail external validation, so this result should be treated as a well-framed hypothesis rather than a ready-to-use tool.

If confirmed in multicenter, prospective studies, the benefit would be concrete: deciding with more information whether to intensify treatment, monitor more closely or consider clinical trials, sparing patients months of toxic chemotherapy with no response. That is the underlying direction of AI-assisted medicine: moving from treating the average to treating the individual patient. It is a small step on the road to precision oncology, but it is in the right direction.