If you've ever tried to figure out how much liver a surgeon can safely remove, you know how frustrating hand-drawing eight squishy liver segments on CT scans is. This paper fixes that.

A team led by Tejas Sudharshan Mathai built a deep learning system that automatically outlines all eight Couinaud liver segments plus the spleen on CT, specifically to help with future liver remnant, or FLR, volumetry [1]. Translation: before a big liver surgery, doctors want to know exactly how much liver will be left standing afterward. That is not a great moment for guesswork, eyeballing, or "close enough."

And the annoying part is that the liver is not organized like a nice IKEA bookshelf. Surgeons think in Couinaud segments, which divide the liver into eight functional regions based on blood vessels and drainage, because those boundaries matter for what can be cut and what absolutely should not be cut [2]. Humans can label these regions on scans, but it is slow, tedious, and about as glamorous as tracing coastlines with a dull pencil.

The liver map problem

This paper used a 3D nnU-Net model, which is basically one of medical imaging's most reliable workhorses - the Honda Civic of segmentation models, except this one stares at CT scans instead of surviving college parking lots. The training set included 498 patients from a public dataset plus one institution, and the model was then tested internally and externally across 311 additional patients, including people with fibrosis, cirrhosis, and portal vein embolization history [1].

That matters because medical AI loves to look heroic on the dataset it grew up with and then immediately get stage fright elsewhere. External validation is where a lot of models discover they were valedictorian of a very small homeschool.

Here, the whole-liver performance was strong in both internal and outside datasets, with Dice scores around 0.98 and low Hausdorff distance errors, which suggests the contours lined up closely with expert labels [1]. That does not mean the model is perfect. It does mean it is getting very good at a job that normally burns radiologist time like a space heater burns electricity in January.

Why surgeons care about leftover liver with near-religious intensity

FLR volumetry is not just bean-counting for organs. If too little healthy liver remains after surgery, the patient can end up with postoperative hepatic insufficiency, which is exactly as bad as it sounds [3]. Modern liver surgery already leans heavily on volumetry, portal vein embolization, and careful pre-op planning to reduce that risk [3].

So a tool that automatically segments the liver into the same functional units surgeons already use could make planning faster, more standardized, and easier to repeat across scans. That is especially useful when patients have distorted anatomy from tumors, cirrhosis, prior treatment, or all the usual ways the human body declines to be geometrically convenient.

This also fits a broader trend in radiology. AI is getting pretty good at the repetitive, high-friction tasks humans hate but still must do carefully: contouring, measuring, triaging, re-checking. RSNA coverage in 2024 made basically this exact point: there is huge interest in AI image analysis, but clinical adoption still lags because tools need to generalize, integrate cleanly, and earn trust in real workflows, not just on a poster session screen [4].

What is actually new here?

Automatic liver segmentation is not new. Automatic Couinaud segmentation that is good enough for surgical planning is the trickier bit.

Earlier work from some of the same research orbit showed AI could estimate liver segmental volume ratios and spleen volume for cirrhosis assessment on CT [5]. Other recent studies have automated future liver remnant assessment by segmenting vessels and modeling blood-free FLR before hepatectomy [6]. A 2026 paper also reported automatic Couinaud segmentation using a hierarchical deep learning network [7].

What makes this new paper interesting is the combination of goals: eight Couinaud segments, spleen, CT-based automation, and a direct link to FLR volumetry [1]. In other words, it is not just "look, the pixels are segmented." It is "look, this could slot into a clinical decision that already matters."

And yes, under the hood this sits in the very busy universe of U-Net-style medical segmentation, which remains dominant because it works absurdly well for an absurd number of tasks. Recent reviews still treat U-Net and its descendants as the default heavy lifters of medical image segmentation, even as transformer-based models keep trying to stroll in wearing cooler jackets [8,9].

The part where we do not get carried away

A good contour is not the same thing as a better patient outcome.

This study was retrospective. It evaluated segmentation accuracy, not whether surgeons changed plans, operated faster, or improved complication rates because of this tool [1]. Also, liver anatomy can get weird in the wild. Scanners differ. Contrast timing differs. Hospitals differ. Humans, very inconsiderately, keep refusing to be standardized.

So the next hurdle is not "can the model draw nice boundaries?" It is "does this save time, reduce variability, and help decisions across messy real-world practice?" That is the grown-up question. The boring question. Also the one that matters.

Still, this is exactly the kind of boring-good AI medicine needs more of. Not chatbot theater. Not sci-fi cosplay. Just software that quietly handles the grindy parts of care so specialists can spend more time on judgment and less time playing anatomy tracing simulator.

References

1. Mathai TS, Balamuralikrishna PTS, Batheja V, et al. Automated Delineation of Couinaud Segments on CT for Future Liver Remnant Volumetry. Radiology: Artificial Intelligence. 2026. DOI: 10.1148/ryai.250808. PubMed: 42089796

2. Liver segment. Wikipedia. Background on Couinaud liver anatomy. https://en.wikipedia.org/wiki/Liver_segment

3. Haddad A, Lendoire M, Maki H, et al. Liver volumetry and liver-regenerative interventions: history, rationale, and emerging tools. Journal of Gastrointestinal Surgery. 2024;28(5):766-775. DOI: 10.1016/j.gassur.2024.02.020. PubMed: 38519362

4. Henderson M. How AI Is Reshaping Musculoskeletal Imaging. RSNA News. March 28, 2024. https://www.rsna.org/news/2024/march/ai-reshaping-msk-imaging

5. Lee S, Elton DC, Yang AH, et al. Fully Automated and Explainable Liver Segmental Volume Ratio and Spleen Segmentation at CT for Diagnosing Cirrhosis. Radiology: Artificial Intelligence. 2022;4(5). DOI: 10.1148/ryai.210268

6. Xie T, Zhou J, Zhang X, et al. Fully automated assessment of the future liver remnant in a blood-free setting via CT before major hepatectomy via deep learning. Insights into Imaging. 2024;15(1):164. DOI: 10.1186/s13244-024-01724-6. PMCID: PMC11211293

7. Lee S, Han K, Shin H, et al. Automatic liver Couinaud segmentation from computed tomography scans with a gradient-enhanced hierarchical cascade deep learning network. Current Problems in Surgery. 2026;75:101957. DOI: 10.1016/j.cpsurg.2025.101957

8. Azad R, Aghdam EK, Rauland A, et al. Medical Image Segmentation Review: The Success of U-Net. IEEE Transactions on Pattern Analysis and Machine Intelligence. 2024;46(12):10076-10095. DOI: 10.1109/TPAMI.2024.3435571. PubMed: 39167505

9. Yao W, Bai J, Liao W, et al. From CNN to Transformer: A Review of Medical Image Segmentation Models. Journal of Digital Imaging. 2024;37(4):1529-1547. DOI: 10.1007/s10278-024-00981-7. PubMed: 38438696

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.