You have probably had that moment in a clinic where everyone is waiting on one more test, and the clock suddenly feels louder than the room.

For patients with unresectable hepatocellular carcinoma, that wait can matter. The cancer is already serious. Many patients also have cirrhosis, which can raise pressure in the portal vein, the liver’s main incoming blood highway. When that pressure climbs, blood finds detours. Some of those detours become esophageal varices, fragile swollen veins in the esophagus that can bleed badly. Elegant? No. The body is resourceful, but sometimes it has the plumbing instincts of an apartment built in 1890.

The usual way to check for these varices is esophagogastroduodenoscopy, or EGD. That means a camera down the throat. It works, but it is invasive, takes coordination, and can delay cancer therapy. In this new Journal of Hepatology study, Rabasco Meneghetti and colleagues ask a simple, beautiful question: can a routine CT scan already tell us enough to act sooner? Their answer is: maybe, and the “maybe” is worth paying attention to (DOI: 10.1016/j.jhep.2026.01.021, PMID: 41679555).

The Art of Seeing What Is Already There

The team studied 489 patients with unresectable HCC treated with atezolizumab plus bevacizumab across five French centers. That therapy matters because bevacizumab can increase bleeding concerns, so clinicians often want varices assessed before treatment moves forward.

Instead of inventing a new scan, the researchers used arterial-phase contrast-enhanced CT images that patients were already getting. There is a wabi-sabi quality here: no perfect new ritual, no shiny extra machine, just imperfect clinical data being asked to speak more clearly.

Their AI pipeline used HepatoSageCT, a deep learning foundation model. A foundation model is basically an apprentice that has stared at a vast amount of medical imaging until it learns useful visual patterns. Not consciousness. Not medical enlightenment. More like a very patient radiology monk with excellent GPU support.

The model’s score was combined with ordinary clinical and imaging features, especially portosystemic shunts, those collateral blood pathways that signal portal hypertension. On their own, portosystemic shunts identified esophageal varices with an AUROC of 0.78. Combined with HepatoSageCT, performance rose to 0.84.

That number is not magic. It means the model became better at ranking who likely had varices and who likely did not. The more clinically vivid result: their decision algorithm missed 4.2% of varices needing treatment, compared with 8.4% using shunts alone, and missed 0% of large varices in the validation cohort.

Negative Space, Clinical Edition

Japanese aesthetics has a concept called ma, the power of space between things. This paper has its own version: the value may come from what clinicians can avoid.

Avoiding unnecessary endoscopies is not a minor convenience. It means fewer invasive procedures, fewer scheduling delays, less patient discomfort, and faster movement toward cancer treatment. A good model here does not need to replace the clinician. It needs to help decide who truly needs the scope and who may not. Like a bouncer for the endoscopy suite, but with fewer velvet ropes and more survival analysis.

The same model also predicted hepatic decompensation, meaning the liver tipping into events such as bleeding, ascites, or hepatic encephalopathy. In the external validation cohort, HepatoSageCT reached a C-index of 0.73 and a hazard ratio of 3.17 for decompensation. A combined score using ascites, splenomegaly, and HepatoSageCT performed similarly, also with a C-index of 0.73. Patients labeled higher risk by the model also had worse overall survival.

That is the paper’s quiet strength. It is not just asking, “Are varices present?” It is asking, “Is this liver close to losing balance?”

Why This Fits the Larger AI Moment

Medical imaging AI is moving from narrow one-task models toward foundation models that can adapt across clinical questions. A 2024 Nature Machine Intelligence paper showed that self-supervised cancer imaging foundation models can help discover CT-based biomarkers when labeled data are scarce (DOI: 10.1038/s42256-024-00807-9, PMCID: PMC10957482). A 2025 CT foundation model study trained on 148,000 CT scans pushed this idea further for 3D radiology tasks (arXiv:2501.09001). Meanwhile, AI work on esophageal varices is gathering into its own small, intense weather system, including a 2024 systematic review of AI for varices detection and bleeding risk (PMCID: PMC12851050) and EVendo validation in HCC (DOI: 10.1007/s10620-024-08449-y, PMCID: PMC11341647).

The challenge, as always, is the gap between promising retrospective performance and reliable clinical behavior. This study used multicenter external validation, which is good. But it still needs prospective testing, workflow studies, calibration across hospitals, and careful checks for bias. AI in medicine has a habit of looking serene in the paper and then getting flustered when the scanner protocol changes. Very relatable, honestly.

The Ikigai of This Model

The model’s purpose, its ikigai, is not to be impressive. It is to reduce delay, reduce unnecessary procedures, and sharpen risk assessment in a tense clinical window.

If future studies confirm these findings, a routine CT could become more than a cancer staging image. It could also serve as a quiet warning system for portal hypertension and liver fragility. Not loud. Not flashy. Just a little more signal from an image already sitting there, waiting to be read.

References

1. Rabasco Meneghetti A, Campani C, Roux C, et al. Detection of esophageal varices and prediction of hepatic decompensation in unresectable hepatocellular carcinoma using AI. Journal of Hepatology. 2026. DOI: 10.1016/j.jhep.2026.01.021, PMID: 41679555.

2. Pai S, et al. Foundation model for cancer imaging biomarkers. Nature Machine Intelligence. 2024. DOI: 10.1038/s42256-024-00807-9, PMCID: PMC10957482.

3. Pai S, Hadzic I, Bontempi D, et al. Vision Foundation Models for Computed Tomography. 2025. arXiv:2501.09001.

4. Esophageal varices detection and bleeding risk assessment with artificial intelligence: a systematic review. 2024. PMCID: PMC12851050.

5. Yang JO, Chittajallu P, Benhammou JN, et al. Validation of a Machine Learning Algorithm, EVendo, for Predicting Esophageal Varices in Hepatocellular Carcinoma. 2024. DOI: 10.1007/s10620-024-08449-y, PMCID: PMC11341647.

Disclaimer: This blog post is a simplified summary of published research for educational purposes. The accompanying illustration is artistic and does not depict actual model architectures, data, or experimental results. Always refer to the original paper for technical details.