Benchmarking Deep Learning for Future Liver Remnant Segmentation in Colorectal Liver Metastasis
A new open, validated dataset and baseline models for AI-assisted liver surgery planning could help prevent a deadly post-operative complication.
The Thesis
When surgeons remove part of a liver to treat colorectal cancer that has spread there, they must leave enough healthy liver tissue behind — called the future liver remnant, or FLR — or the patient risks fatal liver failure after the operation. Estimating FLR volume accurately today requires tedious manual segmentation by radiologists, and no validated public dataset existed to train AI tools for this task. This paper creates that dataset — 197 carefully corrected CT scan volumes — and establishes the first published benchmarks showing how well current deep learning models can automate the work. The best result, a 'cascaded' nnU-Net approach (where the model first finds the liver, then the tumors, then derives the remnant), achieves a Dice score of 0.767 — meaning roughly 77% overlap between the AI's boundary and the manually drawn one, which is promising but still below clinical-grade thresholds. The work's main contribution is infrastructure: a reproducible benchmark that other researchers can now build on.
Catalyst
The CRLM-CT-Seg dataset was released publicly before this work, but contained noisy, unvalidated annotations that made training reliable models impossible. This paper manually corrected all 197 volumes, creating a clean foundation that did not exist until now. Simultaneously, large pretrained medical segmentation models like STU-Net have recently become publicly available, making it practical to test multiple competitive architectures without starting from scratch.
What's New
Previous automated liver segmentation research focused almost entirely on finding the whole liver or individual tumors in isolation — not the specific surgical volume a patient will retain after a partial resection. The CRLM-CT-Seg dataset existed but its annotations were considered too noisy for benchmarking. This paper corrects those annotations, defines two competing pipeline strategies (cascaded step-by-step versus direct end-to-end segmentation), and tests three modern architectures — nnU-Net (a self-configuring 3D segmentation framework), SwinUNETR (a transformer-based model), and STU-Net (a large pretrained medical segmentation network) — providing the first apples-to-apples comparison on a validated holdout set.
The Counter
A Dice score of 0.767 sounds like progress, but surgical planning demands much higher precision — errors in FLR estimation can directly determine whether a surgeon proceeds with an operation or not, and a 23% boundary mismatch is clinically meaningful. The dataset contains only 197 cases from a single public source, which is small by modern standards and likely not representative of scanner diversity, patient demographics, or the full range of resection types seen across different hospital systems. The paper establishes baselines but does not validate performance against actual surgeon decisions or patient outcomes, so the clinical relevance of these Dice numbers is genuinely unknown. All three tested architectures are general-purpose medical segmentation tools — none was specifically designed for the hepatic anatomy problem — and it is entirely possible that a purpose-built approach with richer surgical metadata would leapfrog these results quickly. Finally, the 'first benchmark' framing, while accurate, also means there is no external validation to confirm these numbers will hold on independent institutional data.
Longs
- ISRG (Intuitive Surgical) — AI surgical planning as a natural extension of robot-assisted hepatic surgery
- VOLV (Volpara Health, ASX-listed) — medical imaging AI workflow analogy in oncology planning
- GEHC (GE HealthCare) — CT imaging platforms where FLR tools would be embedded
- NVEI — not applicable; prefer MIRM or radiology AI pure-plays
- RXRX (Recursion Pharmaceuticals) — adjacent medical imaging AI infrastructure and dataset curation
Shorts
- Manual volumetry software vendors (e.g., MeVis Medical Solutions, Visage Imaging) — if AI automates FLR calculation, demand for click-intensive manual tools shrinks
- Radiology outsourcing firms charging per-volume for surgical planning reads — automation compresses this workflow
Enablers (Picks & Shovels)
- nnU-Net open-source framework (github.com/MIC-DKFZ/nnUNet) — the self-configuring baseline that achieved the best FLR Dice score
- STU-Net pretrained weights — large-scale medical segmentation pretraining that improved tumor robustness
- CRLM-CT-Seg public dataset (now corrected) — the validated benchmark this entire effort depends on
- 3D Slicer / ITK-SNAP — open annotation tools used for manual correction of CT volumes
- MONAI (Medical Open Network for AI) — PyTorch-based framework commonly used to implement SwinUNETR and related models
Private Watchlist
- Radiomics (Onepoint Global) — liver and oncology CT analysis
- Contextflow — medical image search and segmentation tooling
- Seg3D / open-source spinouts from academic surgical planning labs
- Mint Medical — radiology reporting and structured imaging data
Resources
The Paper
Accurate segmentation of the future liver remnant (FLR) is critical for surgical planning in colorectal liver metastases (CRLM) to prevent fatal post-hepatectomy liver failure. However, this segmentation task is technically challenging due to complex resection boundaries, convoluted hepatic vasculature and diffuse metastatic lesions. A primary bottleneck in developing automated AI tools has been the lack of high-fidelity, validated data. We address this gap by manually refining all 197 volumes from the public CRLM-CT-Seg dataset, creating the first open-source, validated benchmark for this task. We then establish the first segmentation baselines, comparing cascaded (Liver->CRLM->FLR) and end-to-end (E2E) strategies using nnU-Net, SwinUNETR, and STU-Net. We find a cascaded nnU-Net achieves the best final FLR segmentation Dice (0.767), while the pretrained STU-Net provides superior CRLM segmentation (0.620 Dice) and is significantly more robust to cascaded errors. This work provides the first validated benchmark and a reproducible framework to accelerate research in AI-assisted surgical planning.