The work, led by Yue Zhang, JunXuan Wu, DeJian Xing and Shumei Zhang of the School of Computer Science and Artificial Intelligence at Northeast Forestry University in Harbin, China, was published on 14 October 2026 as an open-access research article. The team describes MFANet as a missing-phase-tolerant multi-source frequency–anatomy network, a name that hints at its two central ideas. The first is a masked phase fusion strategy, in which the model learns to combine information from whichever CT phases are available for a given patient while gracefully handling the absence of others. The second is a pair of separate decoders, one dedicated to the liver and one to tumors, allowing the network to optimize each task with its own anatomical and textural priorities rather than forcing a single output head to serve two very different structures.

The evaluation was deliberately rigorous. The researchers assembled five public datasets and split them at the patient level into a development pool of 1,029 cases and 260 within-source test cases, ensuring that no patient appeared in both training and testing. They then added an independent external dataset: a four-phase contrast-enhanced CT collection from Chongqing Yubei District People’s Hospital comprising 361 patients, of whom 278 had primary liver cancer. The remaining 83 patients were cancer-free controls, and the authors excluded them from tumor evaluation. Five competing segmentation models were given exactly the same phase information and label supervision as MFANet, so that any performance difference could be attributed to architecture rather than to unequal inputs.

The primary endpoint was the equal-source mean patient-level tumor Dice score across the five within-source test cohorts. The Dice score, a standard metric in medical image segmentation, measures the overlap between the model’s predicted tumor region and the expert-drawn ground truth, ranging from 0 for no overlap to 1 for perfect agreement. MFANet achieved a mean tumor Dice of 80.84 percent, compared with 79.84 percent for STU-Net-B, the strongest of the five comparators. The paired difference of 1.00 percentage point came with a 95 percent confidence interval of 0.38 to 1.59, which the researchers assessed using a source-stratified bootstrap. That statistical approach resamples the data within each source dataset, protecting the result from being driven by any single cohort. The interval lying entirely above zero indicates a consistent, if modest, advantage for MFANet.

Small lesions are where liver tumor segmentation is hardest and where the clinical stakes are arguably highest, because tiny nodules are easy for both radiologists and algorithms to overlook. In an exploratory analysis that required a lesion-level Dice of at least 0.10 to count a lesion as detected, MFANet reached a sensitivity of 61.2 percent for tumors smaller than 10 millimeters, against 56.5 percent for STU-Net-B. That gap of nearly five percentage points on sub-centimeter lesions suggests the architecture’s frequency-and-anatomy design may be particularly helpful for the fine-grained texture and shape cues that distinguish small tumors from surrounding liver tissue, although the authors characterize this analysis as exploratory rather than a definitive claim.

To probe how much each imaging phase actually matters, the team ran an ablation in which the portal-venous phase was removed from the input. On the same 168-case subset, Dice scores dropped for all six models tested. Notably, the researchers did not establish a between-model difference in the size of that decline, meaning they could not show that MFANet degrades more gracefully than its competitors when this diagnostically rich phase is absent. The result underscores a practical reality: contrast-enhanced phases, especially the portal-venous phase in which liver tumors typically stand out most clearly, carry substantial information that no current architecture can fully reconstruct from a non-contrast scan alone.

External validation provided a second, harder test. On the independent primary liver cancer contrast-enhanced CT dataset, MFANet achieved a tumor Dice of 77.10 percent among the 278 cancer cases, compared with 75.92 percent for STU-Net-B. Performance on data from an entirely different hospital, scanner fleet and imaging protocol is usually lower than on internal test sets, and the roughly four-point drop from the within-source average here is consistent with that pattern. Still, the fact that MFANet retained its edge over the strongest comparator on out-of-distribution data strengthens the case that the improvement is not an artifact of overfitting to the public benchmarks.

The authors are careful about the limits of their findings, and their candor is worth emphasizing. They conclude that MFANet modestly improved tumor segmentation under matched inputs and supervision in the tested cohorts, but they state plainly that its clinical value and its performance on naturally incomplete external scans and on cancer-free controls remain unknown. The external dataset, while independent, was curated for the study; real-world hospital archives contain scans with unpredictable combinations of phases, variable reconstruction kernels and artifacts that no public benchmark fully reproduces. The exclusion of the 83 control patients from tumor evaluation also means the study says nothing about how such a model behaves in patients without tumors, a scenario that matters for screening applications where false positives carry real costs.

The study’s methodology reflects a broader shift in how medical imaging AI is evaluated. Rather than reporting a single headline number on one benchmark, the team used patient-level partitions across five heterogeneous sources, matched inputs for all comparators, a prespecified primary endpoint, and bootstrap confidence intervals stratified by source. Secondary metrics such as the 95th-percentile Hausdorff distance and normalized surface Dice, which capture how far predicted boundaries stray from true ones, complement the overlap-based Dice score. The authors also report expected calibration error, a measure of how well the model’s confidence aligns with its actual accuracy, an increasingly important consideration for tools intended to support clinical decisions.

For clinicians and researchers, the message is twofold. Architectural choices that explicitly model phase availability and separate organ-level from lesion-level decoding can yield measurable gains on liver tumor segmentation, even under fair, tightly controlled comparisons. At the same time, the modest size of those gains, roughly one percentage point of Dice in most comparisons, is a reminder that the field is in a phase of incremental refinement rather than dramatic leaps. The work was supported by the Heilongjiang Province Postdoctoral Fund, the National Natural Science Foundation of China and the Natural Science Foundation of Heilongjiang Province, and the authors declare no competing interests. As hospitals continue to accumulate multiphase CT archives with inconsistent protocols, models like MFANet point toward segmentation tools that meet clinical data as it actually exists: incomplete, heterogeneous and stubbornly variable from one patient to the next.

Subject of Research: Deep-learning segmentation of liver tumors from incomplete multiphase CT scans

Article Title: Mfanet for liver and tumor segmentation from incomplete multiphase CT: a multi-dataset study

Article References: Zhang, Y., Wu, J., Xing, D., & Zhang, S. (2026). Mfanet for liver and tumor segmentation from incomplete multiphase CT: a multi-dataset study. BMC Medical Imaging. https://doi.org/10.1186/s12880-026-02935-7

Image Credits: AI Generated

DOI: 10.1186/s12880-026-02935-7

Keywords: liver tumor segmentation, MFANet, multiphase CT, deep learning, medical imaging, Dice score, hepatocellular carcinoma, missing phase tolerance, STU-Net, contrast-enhanced CT, artificial intelligence, BMC Medical Imaging

The study, published in BMC Medical Imaging, was conducted by Hee Chang Lee, Woo Hyun Kim, Giltae Song, Jin Kyu Gahm, and colleagues at Pusan National University and Pusan National University Hospital. Their goal was to overcome two persistent limitations of prior CT-based artificial intelligence approaches to cholangiocarcinoma. First, most previous studies relied on single-phase images, typically the portal venous phase, discarding the rich temporal information contained in the arterial, portal venous, and delayed phases of a contrast-enhanced examination. Second, earlier models generally depended on expert-defined two-dimensional regions of interest, in which a radiologist manually outlines the suspicious lesion. That manual step not only consumes specialist time but also constrains the algorithm to a flat slice of anatomy, preventing it from learning the true three-dimensional geometry of the biliary system, which winds through the liver and spans both intrahepatic and extrahepatic ducts.

The technical architecture of the new system unfolds in two stages. In the first stage, the researchers applied automated segmentation networks to identify the liver and gallbladder on each CT scan, and from these structures they extracted three-dimensional hepatobiliary patches, volumetric blocks of image data that cover the intrahepatic and extrahepatic bile ducts without any human outlining. This automated patch extraction is what makes the pipeline fully self-contained: given a raw multiphase CT, the system can localize the relevant anatomy and prepare the model inputs without radiologist interaction. The segmentation step was validated with standard overlap and boundary metrics, including the Dice similarity coefficient, Hausdorff distance, and normalized surface dice, ensuring that the extracted volumes reliably encompassed the biliary structures of interest.

The second and most distinctive stage is the supervised contrastive pre-training. In this paradigm, the network is not initially taught to output a diagnosis. Instead, it learns an internal representation of the images by comparing pairs of examples. Crucially, the researchers constructed positive pairs from the same patient: the arterial-phase patch and the delayed-phase patch of one individual were presented to the model as two views of the same underlying case. The contrastive objective pulls together the learned feature embeddings of these paired patches while pushing apart the embeddings of patches from different patients, and the supervision signal ensures that examples sharing the same diagnostic label cluster together in the embedding space. By training across phases in this way, the network is forced to discover features that are consistent within a patient yet discriminative between disease states, effectively learning how cancer and benign disease manifest across the dynamic enhancement patterns of the liver and bile ducts rather than in a single static snapshot.

After this pre-training, the model was fine-tuned as a classifier on portal venous phase patches, the phase in which many hepatobiliary lesions are conventionally evaluated. The retrospective evaluation enrolled 427 patients: 129 with cholangiocarcinoma, 179 with benign gallbladder disease, and 119 with benign bile duct disease. Performance was assessed with stratified five-fold cross-validation, a rigorous scheme in which the data are divided into five partitions and the model is trained and tested five separate times so that every patient appears in a test set exactly once. The researchers reported sensitivity, specificity, balanced accuracy, and the area under the receiver operating characteristic curve, summarizing results across the five folds to capture the stability of the system rather than a single lucky run.

The results demonstrate meaningful diagnostic power. For the task of distinguishing cholangiocarcinoma from benign gallbladder disease, the model achieved a balanced accuracy of 82.43 percent with an area under the curve of 89.79 percent. Differentiating cancer from benign bile duct disease proved harder, as expected given the closer imaging overlap, but the system still reached a balanced accuracy of 69.90 percent and an area under the curve of 78.26 percent. In both tasks, the proposed phase-aware contrastive model achieved a significantly higher pooled area under the curve than every single-phase model it was compared against, with all differences passing Holm-adjusted statistical significance thresholds. It also consistently exceeded a multiphase early-fusion baseline, which simply concatenates the phases as input channels, across five repeated cross-validation runs, although that particular difference did not reach statistical significance in the primary analysis. The pattern of results suggests that the benefit comes not merely from seeing more phases but from learning how the phases relate to one another.

To understand what the network was actually looking at, the team applied Grad-CAM, a visualization technique that highlights the image regions most influential in a model’s prediction. The resulting maps localized prediction-relevant features to periductal regions and intrahepatic parenchymal areas, precisely the zones where cholangiocarcinoma produces its characteristic changes, such as ductal wall thickening, periductal infiltration, and altered parenchymal enhancement. This anatomical plausibility matters for clinical trust: an algorithm whose attention falls on irrelevant structures would be difficult to deploy, whereas one that concentrates on the ducts and surrounding liver tissue behaves in a way that radiologists can inspect and reason about. Interpretability tools of this kind are increasingly regarded as essential for moving medical artificial intelligence from benchmark performance to bedside utility.

The clinical implications are considerable. Because the pipeline is fully automated, it could in principle run in the background of a radiology workflow, flagging CT studies whose features resemble cholangiocarcinoma and prompting closer review, endoscopic ultrasound, or magnetic resonance imaging follow-up. The authors position the system as a complementary decision-support tool rather than a replacement for expert interpretation, with the aim of improving diagnostic consistency, particularly for the challenging perihilar cases where human readers most often disagree. In an era when imaging volumes are rising and subspecialty expertise is unevenly distributed, a second reader that never tires and always applies the same learned criteria could reduce the variability that currently plagues biliary tumor diagnosis. The researchers also note the potential for integration into clinical workflows, a step that would require prospective validation on external datasets and regulatory evaluation.

The study was approved by the Institutional Review Board of Pusan National University Hospital and conducted in accordance with the Declaration of Helsinki, with written informed consent obtained from all participants. The work was funded by the Institute of Information and Communications Technology Planning and Evaluation through grants from the Korean Ministry of Science and ICT, by the National R&D Program for Cancer Control of the Ministry of Health and Welfare, by a Biomedical Research Institute grant from Pusan National University Hospital, and by the Seegene Medical Foundation. The authors declared no competing interests, and the article is published open access under a Creative Commons Attribution 4.0 license.

Looking forward, the approach illustrates a broader trend in medical imaging artificial intelligence: the shift from hand-crafted regions of interest and single-snapshot analysis toward self-supervised and contrastively learned representations that exploit the full richness of modern imaging protocols. Multiphase CT is already acquired routinely in hepatobiliary workups, yet much of its temporal information goes unused by conventional pipelines. By treating different contrast phases of the same patient as naturally paired views, supervised contrastive learning turns routine clinical data into a training signal that requires no additional annotation burden. If validated prospectively, this strategy could extend beyond cholangiocarcinoma to other cancers whose diagnosis hinges on subtle, phase-dependent enhancement patterns, offering a template for building diagnostic systems that see the way contrast itself reveals disease.

Subject of Research: Deep-learning differentiation of cholangiocarcinoma from benign hepatobiliary disease using multiphase CT and supervised contrastive learning

Article Title: Multiphase CT-based differentiation of cholangiocarcinoma using supervised contrastive learning

Article References: Lee, H. C., Kim, W. H., Park, S. H., Lee, H. J., Lee, J. H., Han, S. Y., Song, G., Kim, D. U., & Gahm, J. K. (2026). Multiphase CT-based differentiation of cholangiocarcinoma using supervised contrastive learning. BMC Medical Imaging. https://doi.org/10.1186/s12880-026-02835-w

Image Credits: AI Generated

DOI: 10.1186/s12880-026-02835-w

Keywords: cholangiocarcinoma, bile duct cancer, multiphase CT, supervised contrastive learning, deep learning, medical imaging, computer-aided diagnosis, biliary tract disease, radiology, machine learning, Grad-CAM, Pusan National University