Published online Jan 14, 2027. doi: 10.3748/wjg.125661
Revised: August 5, 2026
Accepted: September 17, 2026
Published online: January 14, 2027
Processing time: 121 Days and 16.6 Hours
Magnetic resonance cholangiopancreatography (MRCP) depicts the biliary tree three-dimensionally but is used mainly for qualitative diagnosis rather than quantitative procedural planning. Endoscopic drainage of biliary strictures, parti
To develop and evaluate a proof-of-concept AI workflow that converts MRCP-derived biliary anatomy into a quantitative 3D stricture model.
This single-center study analyzed 384 MRCP examinations (242 controls and 142 with suspected biliary stricture). Expert-corrected masks and 3D annotations were used to train or evaluate modules for region-of-interest localization, bile duct segmentation, projection-based multi-view classification, synthetic morphology restoration, and centerline-guided tubular mesh generation. Binary stricture detection and three-class severity classification were assessed with patient-level cross-validation after excluding unavailable or task-incompatible labels. Segmentation and restoration were evaluated using Dice scores.
In 380 evaluable cases, projection-based binary classification achieved an accuracy of 0.9770, balanced accuracy of 0.9721, macro F1 score of 0.9749, and receiver operating characteristic area under the curve of 0.9953. Three-class severity classification was less accurate (accuracy 0.8711), indicating that severity stratification is more difficult than binary detection. The best Dice score was 0.8113 in the segmentation comparison and 0.8944 in the extended normal-group experiment. In synthetic restoration, DiceFocalLoss produced the highest Dice score (0.7531), and lesion weighting improved reconstruction of interrupted duct segments. In anatomically simple cases, restored geometry was converted into a centerline-guided tubular mesh, whereas complex hilar or intrahepatic branching remained vulnerable to projection overlap.
MRCP-based AI modeling of biliary strictures is feasible for quantitative 3D assessment and exploratory stent-planning geometry. External validation and prospective clinical evaluation are required before procedural implementation.
Core Tip: This proof-of-concept study shows that magnetic resonance cholangiopancreatography-derived biliary anatomy can be transformed by an integrated artificial-intelligence workflow into a quantitative three-dimensional model of biliary strictures. By combining region-of-interest localization, transformer-based segmentation, projection-based multi-view classification, synthetic morphology restoration, and centerline-guided mesh generation, the pipeline reframes biliary stricture analysis from qualitative image review toward reproducible structural modeling and candidate stent-planning geometry. Binary stricture detection was highly accurate internally, whereas reconstruction of complex hilar or intrahepatic anatomy remained challenging and requires external validation before clinical use.
- Citation: Chun JH, Hong JW, Hou JU, Lee KJ, Park DH, Cha HW, Park SW, Min SJ, Yu W. Artificial intelligence-driven three-dimensional modeling of biliary strictures using magnetic resonance cholangiopancreatography for quantitative assessment and stent planning. World J Gastroenterol 2027; 33(2): 125661
- URL: https://www.wjgnet.com/1007-9327/full/v33/i2/125661.htm
- DOI: https://dx.doi.org/10.3748/wjg.125661
Biliary strictures are clinically important causes of obstructive cholestasis because management requires both reliable etiologic assessment and effective restoration of bile flow[1]. Contemporary guidelines emphasize two complementary priorities: Confirming or excluding malignancy and achieving adequate drainage, with diagnostic and therapeutic strategies differing by stricture location, including extrahepatic, perihilar, and intrahepatic disease[2,3]. Cross-sectional imaging and magnetic resonance cholangiopancreatography (MRCP) are therefore central before endoscopic intervention because they delineate the ductal anatomy, the level of obstruction, the extent of ductal involvement, and the relationship between the stricture and upstream branches[4]. Because hepatobiliary treatment selection remains strongly anatomy-dependent and outcome-sensitive – the level and extent of ductal involvement determine whether unilateral or bilateral drainage is required and how much functional liver can be decompressed – an accurate preprocedural appreciation of ductal geometry is directly linked to drainage success and to the risk of early reintervention[5].
Preprocedural anatomic assessment is especially important for ERCP-guided drainage of hilar and intrahepatic strictures[6]. These lesions often involve branching ducts with patient-specific curvature, segmental dilation, and overlapping ductal courses. Underestimation of lesion length, ductal angulation, or branch involvement may result in incomplete drainage, suboptimal stent length or diameter, inadequate coverage of the stenotic segment, or early need for reintervention[7]. However, the clinical information obtained from MRCP is usually summarized qualitatively, while ERCP fluoroscopy collapses the three-dimensional (3D) ductal structure into a two-dimensional projection[8]. As a result, the rich volumetric information contained in MRCP is not routinely converted into reproducible measurements of stric
Artificial intelligence (AI) may help bridge this gap by converting diagnostic imaging into quantitative procedural information. Quantitative MRCP research has shown that automated biliary metrics can provide clinically meaningful information in cholangiopathies such as primary sclerosing cholangitis, and recent AI applications in biliopancreatic imaging and endoscopy suggest a growing role for data-driven decision support[9,10]. For biliary stricture planning, however, a useful workflow must do more than classify the presence of obstruction. It should localize the affected duct, preserve the thin and sparsely distributed biliary tree, estimate the morphology of signal-deficient segments, and generate interpretable geometric outputs that can be reviewed by endoscopists or engineers. In particular, by making the branching relationships of the biliary tree explicit before intervention, AI-based anatomical analysis may reduce the likelihood of misidentifying the target duct and thereby support safer preprocedural planning[11].
We therefore designed a multi-stage, proof-of-concept AI workflow using clinical MRCP volumes and expert-corrected biliary masks. The pipeline combines 3D stricture localization, transformer-based bile duct segmentation, projection-based multi-view classification, synthetic restoration of interrupted ductal morphology, and centerline-guided generation of candidate stent-planning geometry. The objective was not to validate an autonomous diagnostic system or an implantable device, but to test whether MRCP-derived biliary anatomy could be transformed into a quantitative 3D model that may support future preprocedural assessment of stricture location, length, curvature, branch involvement, and potential stent coverage. Specifically, the unmet clinical need addressed here is the absence of an automated method that converts a routinely acquired MRCP into reproducible, quantitative 3D measurements – stricture location, length, curvature, and branch involvement – usable for preprocedural biliary drainage and stent planning rather than for diagnosis alone.
This was a single-center proof-of-concept methodological study using clinically acquired MRCP examinations from Hallym University Dongtan Sacred Heart Hospital. The study was designed to evaluate an image-analysis and geometric-modeling workflow for biliary stricture assessment, rather than to classify stricture etiology or to compare clinical drainage strategies. The protocol was reviewed and approved by the Institutional Review Board of Hallym University Dongtan Sacred Heart Hospital (approval No. 2023-09-003-001).
The dataset included 384 MRCP examinations: (1) 242 normal controls; and (2) 142 cases with radiologically and/or clinically suspected biliary stricture. Normal controls had no visible stricture or clinically meaningful biliary obstruction on expert MRCP review. Cases were excluded from individual experiments when labels were missing, undefined, or incompatible with a task-specific definition. Consequently, the number of analyzable samples differed across classification, segmentation, localization, and restoration experiments.
DICOM MRCP studies were converted to NIfTI format to preserve volumetric geometry and enable voxel-wise processing. Ground-truth biliary masks were generated using Siemens Radiomics software and refined by clinical experts. Ground-truth severity was defined on MRCP as an ordinal, clinically derived grade based on the continuity of the high-signal biliary lumen. A grade of 0 indicated no stricture (a normal, continuous duct); grade 1 (partial stricture) indicated a reduced luminal caliber in which the bright luminal signal remained continuous across the lesion, corresponding to restricted but preserved bile flow; and grade 2 (complete stricture) indicated interruption or loss of the luminal signal at the lesion (signal cut-off), corresponding to obstructed bile flow. Grades were assigned by expert consensus using this uniform lumen-continuity criterion, and this same ordinal label served as the reference standard for the three-class severity model.
Seed points were placed along the biliary tree, connected structures were extracted automatically, and preliminary masks were corrected slice by slice to remove non-biliary signal, recover fine branches when visible, and improve anatomic plausibility. Suspected stricture regions were annotated in 3D Slicer as bounding boxes and converted into model-compatible labels. The study included complementary modules for lesion localization, bile duct segmentation, projection-based multi-view classification, synthetic stricture generation, morphology restoration, 3D reassembly, and candidate stent-planning geometry generation. Figure 1 summarizes the two projection-based workflows used for morphology restoration and multi-view classification, whereas 3D stricture localization is shown separately in Figure 2.
The ground-truth biliary masks and the 3D stricture bounding boxes were independently reviewed and corrected against the source MRCP by two radiologists (Min SJ and Yu W) experienced in hepatobiliary MRCP interpretation; annotation discrepancies were resolved by consensus, and interobserver agreement was not formally quantified in this proof-of-concept study.
Data provenance was kept distinct across tasks. Clinical masks and manually defined 3D bounding boxes served as reference annotations for localization, segmentation, and classification, whereas synthetically interrupted ducts were used only to create paired inputs and targets for morphology-restoration training.
Stenotic regions were localized using MedYOLO, a 3D object-detection framework for volumetric medical images[12]. Because biliary strictures are small, sparse, and often elongated, the detector was used as a region-of-interest localization tool rather than as a standalone diagnostic classifier. Input resolution was increased relative to standard settings to reduce loss of small ductal features. The anchor-matching rule was relaxed to accommodate narrow targets, and confidence and intersection-over-union thresholds were tuned empirically during development. Predicted boxes were reviewed qualitatively to confirm whether they included the affected ductal region, even when they were broader than the manually annotated stenosis (Figure 2).
Localization performance was quantified using precision, recall, mean average precision (mAP) at an intersection-over-union threshold of 0.50, and mAP averaged over thresholds from 0.50 to 0.95. Because small target boxes and broader predictions can produce low overlap despite spatial proximity, matched predictions were additionally evaluated using the 3D Euclidean L2 distance between predicted and ground-truth box centers and the absolute center error along each spatial axis. Distances were calculated in physical coordinates and reported in millimeters with 95%CIs.
Bile duct segmentation was performed using Swin-UNETR, a transformer-based 3D segmentation architecture that can model sparse tubular structures with long-range context[13-15]. This architecture was selected because the biliary tree occupies a small fraction of the MRCP volume and because MRCP signal loss at the stenotic segment may make the duct appear discontinuous. Intensities were normalized to the range of -175 to 250 and rescaled to 0-1. Hyperparameter tuning varied input size, window size, and training duration while maintaining a consistent optimization framework. To encourage preservation of ductal continuity near the lesion, the stricture region was assigned a distinct label and a higher relative weight during loss calculation.
For classification, each MRCP volume was projected onto the XY, YZ, and ZX planes and rotated at 10-degree intervals, producing 108 projected images per case. Each two-dimensional input contained the grayscale projection, projected bile duct mask, and masked region-of-interest image. Plane identity and rotation angle were encoded as metadata using one-hot plane indicators and sine-cosine angular features. Image features were extracted with an ImageNet-pretrained SwinV2-Base backbone and fused with metadata in a late-fusion classification head[16]. Binary stricture classification and three-class severity classification were evaluated with four-fold stratified cross-validation at the case level, with an additional internal 20% hold-out ensemble using soft-voting aggregation. One undefined-label case and three intrahepatic cases were excluded from classification, leaving 380 evaluable samples.
In the late-fusion framework, projection metadata (plane orientation and angle, encoded as a 5-dimensional vector) were mapped to a 32-dimensional embedding and concatenated with SwinV2-derived image features (1024-dimensional) prior to classification. Case-level predictions were derived by uniformly averaging softmax probabilities across all projected views. Model uncertainty was quantified via 10000 patient-level bootstrap resamples (clustering views by patient for view-level analyses), and multiclass receiver operating characteristic area under the curve (ROC AUC) was evaluated using a macro-averaged one-vs-rest formulation.
Because paired pre-stricture and post-stricture MRCP scans were unavailable, synthetic strictures were generated from normal bile duct masks to provide surrogate supervision[15]. The common bile duct was used as the primary source structure because it is relatively continuous and suitable for controlled deformation. Local ductal signals were removed or attenuated to simulate interruption or narrowing, and the target for restoration was the corresponding unaltered ductal segment[16]. The 3D biliary volumes were projected onto orthogonal two-dimensional planes, restored with a Stable Diffusion v1.5-based inpainting strategy, and back-projected into 3D space by retaining regions consistently reconstructed across projections[17]. Restoration experiments compared lesion weighting, training duration, and imbalance-aware loss functions, including DiceFocalLoss, Tversky loss, generalized Dice loss, weighted Dice cross-entropy, and Top-K% loss (Supplementary Tables 1-3)[18]. Representative outputs were used to assess whether restored morphology was anatomically plausible or inconsistent (Figure 3 and Supplementary Figure 1)[19].
For each selected normal, expert-corrected bile-duct mask, the intact common bile duct segment served as the restoration target. A paired synthetic input was created by locally masking or attenuating the corresponding segment to reproduce a controlled interruption or narrowing. This procedure simulated loss of ductal continuity but did not attempt to synthesize the tissue-level appearance of tumor infiltration, inflammation, stone impaction, postoperative change, or stent artifact.
To complement volumetric Dice in the tubular morphology analysis, we additionally measured the 95% Hausdorff distance, average symmetric surface distance, centerline Dice (clDice), and mean centerline distance. The corresponding 95%CIs are reported in Supplementary Table 4.
During 3D back-projection, the number of projection views supporting each reconstructed voxel was retained and visualized as a multi-view projection-support map. Regions supported by fewer views, isolated reconstructed compon
The restored stricture-supporting region was converted into a centerline-guided candidate stent-planning mesh. A provisional centerline was extracted using a skeleton-graph approach, resampled, smoothed, and constrained by the nearest duct component when restoration was noisy or fragmented. A regularized radius profile was sampled along the centerline and converted into a tubular triangular mesh in physical coordinates. Minimum-diameter and reference-ratio constraints were applied to maintain a manufacturable geometry while respecting the surrounding duct caliber. This output was treated strictly as a research-use planning geometry, not as an implant-ready device. Translation toward clinical use would require mechanical testing, biocompatibility and sterility validation, regulatory review, and prospective clinical evaluation.
The full dataset consisted of 384 MRCP examinations, including 242 normal controls and 142 cases with biliary stricture. Of the 384 evaluable examinations, 380 were entered into the projection-based classification analysis. Four case-arm examinations were excluded: (1) One with an undefined binary ground-truth label; and (2) Three intrahepatic cases whose peripheral ductal disease was not reliably captured by the whole-tree multi-view projection. The excluded intrahepatic cases represented three of the seven intrahepatic strictures in the cohort (Table 1); the remaining four intrahepatic and all 29 perihilar strictures were retained in the classification set. Localization, segmentation, and morphology-restoration experiments, as well as the baseline characterization, used the full cohort. The per-module inclusion and exclusion of examinations across all analytical modules is summarized in Supplementary Figure 2. The workflow generated complementary outputs for region-of-interest localization, biliary segmentation, stricture classification, morphology restoration, and candidate stent-planning geometry (Figure 1). Because this was an internal proof-of-concept study, model performance values should be interpreted as descriptive internal results rather than externally validated diagnostic accuracy.
| Characteristic | Controls (n = 242) | Cases (n = 142) | Total (n = 384) |
| Age (years) | 48.8 ± 14.1 | 68.0 ± 11.2 | 55.9 ± 16.0 |
| Sex | |||
| Male | 121 (50.0) | 81 (57.0) | 202 (52.6) |
| Female | 121 (50.0) | 61 (43.0) | 182 (47.4) |
| Stricture characteristics (cases only, n = 142) | |||
| Ground-truth severity | |||
| Grade 0 (none) | 242 (100) | 0 (0) | 242 (63.0) |
| Grade 1 (mild) | 0 (0) | 77 (54.2) | 77 (20.1) |
| Grade 2 (severe) | 0 (0) | 65 (45.8) | 65 (16.9) |
| Etiology | |||
| Malignant | 129 (90.8) | ||
| Benign | 11 (7.7) | ||
| Indeterminate | 2 (1.4) | ||
| Distribution | |||
| Focal | 137 (96.5) | ||
| Diffuse | 5 (3.5) | ||
| Location | |||
| Extrahepatic | 106 (74.6) | ||
| Intrahepatic | 7 (4.9) | ||
| Perihilar/hilar | 29 (20.4) | ||
| Bismuth type | |||
| Bismuth type I | 4 (2.8) | ||
| Bismuth type II | 3 (2.1) | ||
| Bismuth type IIIa | 7 (4.9) | ||
| Bismuth type IIIb | 2 (1.4) | ||
| Bismuth type IV | 5 (3.5) | ||
| Bismuth type, indeterminate | 8 (5.6) | ||
| Stricture length, mm, median (range) | 25.0 (6.7-120.0) | ||
| Max upstream duct diameter, mm | 13.7 ± 3.0 | ||
The baseline demographic and clinical characteristics of the cohort are summarized in Table 1. The mean age was 55.9 ± 16.0 years (controls 48.8 ± 14.1 years; cases 68.0 ± 11.2 years), and 202 patients (52.6%) were male. Among the 142 stricture cases, 129 (90.8%) were malignant, 11 (7.7%) benign, and 2 (1.4%) indeterminate; by location, 106 (74.6%) were extrahepatic, 29 (20.4%) perihilar (hilar), and 7 (4.9%) intrahepatic, and 137 (96.5%) were focal. The median stricture length was 25.0 mm (range 6.7-120.0), and the expert-assigned ground-truth severity was grade 1 (mild) in 77 cases (54.2%) and grade 2 (severe) in 65 cases (45.8%).
MedYOLO showed limited exact bounding-box agreement, with precision of 0.197, recall of 0.056, mAP@0.50 of 0.0142, and mAP@0.50:0.95 of 0.0014. Because the target strictures occupied only a small volume and predicted boxes tended to be broader than manual annotations, spatial proximity was additionally assessed among matched predictions. The mean 3D center L2 distance was 10.13 mm (95%CI: 5.55-14.71 mm). Mean absolute X-axis, Y-axis, and Z-axis center errors were 3.17 mm (1.25-5.10 mm), 3.21 mm (0.79-5.63 mm), and 8.40 mm (4.56-12.24 mm), respectively. These findings suggest approximate localization of the anatomical vicinity among matched predictions despite poor agreement with the precise box extent. MedYOLO was therefore interpreted as a coarse region-of-interest proposal module rather than a precise lesion-delineation method.
Swin-UNETR captured the overall morphology of the biliary tree across representative volumes. In the hyperparameter comparison, the best Dice score was 0.8113 with an input size of 100 × 200 × 200 and a window size of 96 × 96 × 96 (Table 2 and Supplementary Table 5). In the extended normal-group experiment, the best Dice score reached 0.8944. These values indicate that the global ductal structure can be segmented with reasonable overlap; however, the stenotic region remained more difficult than non-strictured duct because local signal loss can make the duct appear interrupted. Therefore, global Dice should not be interpreted as sufficient evidence that the clinically critical narrowed segment is always reconstructed accurately.
| Analysis component | Detailed experiment summarized | Representative condition for main table | Primary metric | Result | Main interpretation |
| Bile duct segmentation | Swin-UNETR input resizing and window-size comparison | Input size 100 × 200 × 200; window size 96 × 96 × 96 | Dice score | 0.8113 | Best result in the segmentation hyperparameter comparison; supports feasibility of whole biliary-tree segmentation |
| Bile duct segmentation | Extended normal-group segmentation experiment | Best-performing model in the extended normal-group setting | Dice score | 0.8944 | Higher overlap in non-strictured anatomy; useful for showing that stenotic signal loss is the harder condition |
| Morphology restoration | Loss-function comparison | DiceFocalLoss | Dice score | 0.7531 | Best restoration result; appropriate main-text evidence for imbalance-aware restoration of thin ductal structures |
| Morphology restoration | Training-iteration comparison | 300000 iterations | Dice score | 0.7508 | Longer training improved reconstruction of interrupted duct segments |
| Morphology restoration | Lesion-weighting comparison | Lesion weight = 9 | Dice score | 0.7245 | Targeted lesion weighting improved restoration around the synthetically interrupted segment |
| Three-dimensional reassembly and planning geometry | Representative qualitative reconstruction and centerline-guided mesh generation | Simple extrahepatic anatomy with limited branch overlap | Qualitative feasibility | Plausible restoration/candidate mesh | Shows technical feasibility for support path, approximate caliber, and curvature, but should not be interpreted as a validated stent design |
Projection-based multi-view classification showed strong internal performance for binary stricture detection. View-level out-of-fold analysis achieved an accuracy of 0.9478 and a ROC AUC of 0.9768. Aggregation across 108 views improved case-level performance to an accuracy of 0.9770, balanced accuracy of 0.9721, macro F1 score of 0.9749, and ROC AUC of 0.9953 (Table 3). The internal 20% hold-out ensemble remained robust, with an accuracy of 0.9737 and ROC AUC of 0.9970. Gaussian projection produced the best overall result among the tested projection methods (Supplementary Table 6). Accuracy generally improved as more projection views and plane combinations were used; the all-plane 36-view configuration achieved the highest accuracy in the view-count analysis (accuracy, 0.976) (Supplementary Table 7, Supplementary Figure 3). In contrast, three-class severity classification was less reliable, with case-level accuracy of 0.8711 and ROC AUC of 0.9328, suggesting that severity stratification is a more ambiguous task than binary detection. These values should be read as internal, descriptive performance obtained on a single-center dataset using projected masks and region-of-interest information, and therefore do not, in themselves, establish clinical diagnostic readiness. Detailed classification metrics and their corresponding 95%CIs across all evaluation levels – including view-level, patient-level, hold-out ensemble, and three-class severity analyses – are presented in Table 3.
| Experiment | Evaluation level/aggregation | Accuracy (95%CI) | Balanced accuracy (95%CI) | Macro F1 (95%CI) | ROC AUC (95%CI) |
| Binary presence, view OOF | Individual projected view; patient-cluster bootstrap | 0.9478 (0.9302-0.9636) | 0.9399 (0.9187-0.9590) | 0.9430 (0.9236-0.9603) | 0.9768 (0.9649-0.9866) |
| Binary presence, patient OOF | Case-level aggregation across projected views | 0.9770 (0.9605-0.9934) | 0.9721 (0.9494-0.9909) | 0.9749 (0.9564-0.9928) | 0.9953 (0.9903-0.9988) |
| Binary presence, hold-out ensemble | Internal 20% hold-out ensemble; soft voting | 0.9737 (0.9342-1.0000) | 0.9643 (0.9107-1.0000) | 0.9713 (0.9262-1.0000) | 0.9970 (0.9888-1.0000) |
| 3-class severity, patient OOF | Case-level severity classification | 0.8711 (0.8395-0.9000) | 0.7833 (0.7302-0.8355) | 0.7817 (0.7287-0.8330) | 0.9328 (0.9070-0.9557) |
Synthetic restoration experiments showed that targeted weighting and longer training improved reconstruction of the interrupted duct segment. Dice scores increased from 0.5691 to 0.7245 as lesion weight increased from 3 to 9 (Table 2, Supplementary Table 1), and from 0.6957 to 0.7508 as training increased from 30000 iterations to 300000 iterations (Table 2, Supplementary Table 2). Among loss functions, DiceFocalLoss achieved the highest Dice score (0.7531), followed by Tversky loss (0.6880), weighted Dice cross-entropy (0.6532), Top-K% loss (0.5796), and generalized Dice loss (0.4537) (Table 2, Supplementary Table 3). In the representative qualitative comparison, Swin-UNETR better preserved elongated and partially disconnected biliary structures than MedNeXt (Supplementary Figure 4). Distance-based and centerline-based evaluations showed that volumetric overlap and geometric fidelity did not always rank loss functions identically (Supplementary Table 4). For example, although weighted Dice cross-entropy achieved the highest clDice, it exhibited larger surface and centerline distance errors than the center-loss setting. This discrepancy underscores the necessity of evaluating overlap and distance-based metrics in tandem rather than relying on a single measure.
The two-dimensional inpainting and back-projection strategy produced plausible 3D restoration in simple extrahepatic anatomy with limited branch overlap (Figure 3 and Video 1). Failures occurred in complex hilar or intrahepatic branching, where overlapping projections caused incorrect connections or hallucinated duct-like structures. These findings indicate that the current restoration module is most reliable for relatively simple anatomy and should be considered a morphology-support tool rather than a definitive reconstruction method.
Cross-view consistency was visualized using a projection-support map to flag low-support or implausible regions for manual review (Supplementary Figure 5). Representative failure cases exhibited discontinuous restoration and false-positive fragments in overlapping branches (Supplementary Figure 6), emphasizing that all reconstructed outputs represent model-generated hypotheses rather than observed anatomy.
The restored stricture-supporting region was converted into a centerline-guided tubular mesh. The generated mesh followed the restored duct course and covered the narrowed segment when overlaid on the original stenotic duct (Supplementary Figure 1). This result demonstrates technical feasibility for generating candidate support length, approximate caliber, and curvature in a patient-specific coordinate system. However, the output remains a computational planning model and should not be interpreted as a validated stent design. We stress that this mesh is a research-use computational planning geometry, not a validated or implant-ready stent design.
This proof-of-concept study demonstrates that MRCP-derived biliary anatomy can be processed through an integrated AI workflow for stricture localization, bile duct segmentation, multi-view classification, morphology restoration, and candidate stent-planning geometry. The main contribution is not a single classification metric, but the demonstration of complementary projection-based and 3D modeling modules that transform MRCP-derived biliary anatomy into interpretable outputs for stricture assessment and exploratory planning. This reframes biliary stricture analysis from qualitative image review toward quantitative structural modeling that may eventually support preprocedural planning. We emphasize that the three principal outputs of the workflow – stricture classification, 3D morphological reconstruction, and candidate planning geometry – are related but not equivalent and should not be conflated: (1) Classification answers whether a stricture is present or how severe it is and is judged by discrimination metrics such as accuracy and area under the receiver operating characteristic curve; (2) Reconstruction attempts to recover the true ductal morphology at the lesion and must be judged by geometric fidelity rather than by classification accuracy; and (3) Planning geometry converts a reconstruction into a manufacturable candidate device path whose adequacy depends on mechanical and procedural criteria that neither earlier stage evaluates. Accordingly, a high binary-classification performance does not by itself imply faithful reconstruction, and a plausible reconstruction does not guarantee a clinically deployable stent-planning geometry, so each stage requires its own dedicated validation.
The clinical rationale is consistent with current guideline-based practice, in which biliary stricture management depends on both diagnosis and drainage strategy[1]. MRCP provides volumetric information before ERCP, but conventional fluoroscopy and routine reporting do not directly translate this information into measurements that are useful for selecting stent length, estimating the coverage range, or anticipating branch involvement. Prior work on MRCP-based 3D reconstruction in malignant hilar stricture has suggested that reconstructed biliary anatomy can improve anatomical understanding during endoscopic planning[17]. Our study extends that concept by automating several components of the workflow and by linking reconstruction to a centerline-guided candidate geometry. Binary projection-based classification performed strongly after multi-view aggregation. This suggests that projection ensembles can capture diagnostically relevant patterns such as upstream dilation, ductal interruption, and global changes in biliary morphology. Nevertheless, the high ROC AUC should be interpreted cautiously. The evaluation was internal, labels were derived from a single-center dataset, and the classifier used projected masks and region-of-interest information rather than raw MRCP alone. Therefore, this result supports feasibility for case prioritization or workflow triage, but it should not be presented as clinical diagnostic accuracy until validated on external institutions, acquisition protocols, and etiologic subgroups.
Severity classification was less accurate than binary detection. This is scientifically plausible because severity labels are often ordinal, may depend on clinical context, and can overlap morphologically. A mild stricture in a small duct, a short high-grade extrahepatic stricture, and a complex hilar narrowing may differ in drainage implications even when their two-dimensional projections appear similar. Future work should test clinically grounded labels such as Bismuth-Corlette class, drainage-relevant liver segment involvement, endoscopist-selected target ducts, or measured pre-stenotic and post-stenotic caliber ratios rather than relying only on broad categorical severity labels.
Segmentation and restoration results highlight a key methodological issue: Global overlap metrics may not reflect the clinically most important anatomy. A Dice score near or above 0.8 can still be compatible with local failure at a short stenotic segment because the narrow target occupies few voxels. The improvement obtained with lesion weighting and DiceFocalLoss is consistent with the known challenge of class imbalance in medical image segmentation[18,19]. For biliary applications, future evaluation should therefore include branch-level continuity, centerline-distance error, stenosis-length error, local radius error, and drainage-territory involvement in addition to voxel-wise Dice.
Accordingly, the revised evaluation includes 95% Hausdorff distance and average symmetric surface distance for boundary discrepancy and clDice and mean centerline distance for tubular centerline preservation. These measures provide complementary information but do not directly quantify branch-by-branch continuity, local radius error, absolute stricture-length error, or failure rates in hilar and intrahepatic subgroups. Those clinically specific analyses remain necessary in future external and prospective validation.
The generative restoration module represents the most experimental component of the workflow. Stable Diffusion-based inpainting can plausibly fill interrupted ductal segments in simple anatomy, but it can also produce anatomically incorrect connections when projections overlap or when intrahepatic branches cross in two-dimensional views. This risk is particularly important because hallucinated ductal continuity could lead to an erroneous stent-planning pathway. The method should therefore incorporate uncertainty estimation, branch-aware topology constraints, multi-angle consistency checks, and direct 3D generative approaches before clinical translation. Restored segments should be displayed as model estimates rather than ground truth.
The projection-support map provides a heuristic measure of geometric consistency across projection views rather than a calibrated probability of anatomical correctness. The current framework does not provide formal stochastic uncertainty estimation or topology verification. Future safeguards should therefore include graph-based branch-connectivity analysis, repeated diffusion sampling, persistent-homology-based verification, and calibrated voxel- or ensemble-level uncertainty. Until these methods are validated, reconstructed segments should be interpreted only as candidate regions for expert review and downstream analysis, particularly in complex hilar or intrahepatic anatomy.
The centerline-guided mesh provides a tangible bridge between image analysis and interventional planning. Its potential value lies in estimating a candidate support path, approximate diameter profile, curvature, and length in a patient-specific coordinate system. Similar patient-specific 3D printing and biliary stent prototyping studies have emphasized the importance of anatomical realism, material properties, and mechanical validation for translational use[20-23]. Consistent with those studies, our output should be considered a geometric planning aid only. It does not address radial force, migration risk, deployability through an endoscope, tissue interaction, radiopacity, durability, sterility, or regulatory requirements. Clinically, the centerline-guided geometry could support several preprocedural decisions in a concrete way. First, for stent-length selection, the arc length measured along the reconstructed centerline between the proximal and distal shoulders of the narrowed segment provides a patient-specific estimate of stenosis length, to which standardized proximal and distal landing margins can be added to derive a candidate covered-stent length, replacing visual length estimation from a foreshortened two-dimensional fluoroscopic projection. Second, for drainage-route planning, the reconstructed branch topology makes explicit which segmental or sectoral ducts lie upstream of the stricture and whether the hilar confluence is involved, directly informing the choice between unilateral and bilateral drainage and the selection of the target duct expected to decompress the largest functional liver volume. Third, because the mesh is expressed in physical patient coordinates with an associated radius and curvature profile, it can serve as a geometric substrate for procedural simulation – for example, rehearsing guidewire and catheter trajectory, anticipating the angulation a device must negotiate at the stricture, or producing a 3D-printed phantom for hands-on planning. We reiterate that these applications remain hypothetical and would each require prospective comparison against expert procedural plans and clinical outcomes before influencing real decisions.
This study has several limitations. First, it was conducted at a single center and lacks external validation; therefore, model performance may not generalize to different scanners, MRCP protocols, reconstruction parameters, or patient populations. Second, the dataset was relatively small for training and evaluating deep-learning models and was not designed to evaluate etiologic diagnosis, malignant vs benign strictures, or outcomes after drainage. In addition, the projection-based classification set excluded a small number of anatomically challenging examinations – one undefined-label case and three intrahepatic cases whose peripheral duct was poorly represented by the multi-view projection. Although four intrahepatic and all perihilar strictures were retained, intrahepatic and hilar strictures are among the most demanding settings for both imaging interpretation and drainage, and are under-represented in this single-center cohort. The reported classification performance should therefore be regarded as an internal, potentially optimistic estimate for peripheral intrahepatic disease, and dedicated evaluation in intrahepatic-enriched and hilar-enriched, ideally multicenter, cohorts is required before the approach can be considered robust across the full anatomical spectrum. Third, MedYOLO showed low exact bounding-box agreement (precision, 0.197; recall, 0.056; mAP@0.50, 0.0142; mAP@0.50:0.95, 0.0014), while the center-distance analysis among matched predictions supported only coarse region-of-interest localization rather than precise lesion delineation. Fourth, restoration was trained with synthetic strictures rather than paired real pre-treatment and post-treatment images, and the synthetic lesions may not fully capture tumor infiltration, inflammation, stone-related obstruction, postoperative anatomy, or metal-stent artifacts. Fifth, the current projection-based reconstruction is vulnerable to branch overlap and hallucination in complex hilar or intrahepatic anatomy. Sixth, the stent-planning geometry was evaluated computationally only and was not tested for mechanical performance, manufacturability, procedural usability, or clinical benefit. In addition, because the synthetic strictures were produced by masking segments of the normal common bile duct, they reproduce the geometric interruption of the lumen but not the tissue-level appearance of real disease – tumor infiltration, wall thickening, inflammatory change, stone impaction, or metallic-stent artifact – so the restoration model may behave differently on genuinely pathological ducts, and its transferability to clinical cases remains unproven. Finally, the three-class severity grade is an ordinal, imaging-based reference standard defined by the continuity of the biliary lumen on MRCP (partial stricture with a preserved lumen vs complete stricture with a signal cut-off); although this criterion is reproducible and clinically motivated – a traversable residual lumen is relevant to guidewire negotiation and to the feasibility of transpapillary drainage – it reflects the imaging appearance of obstruction rather than a validated, one-to-one prescription of drainage strategy, and linking these grades to specific interventional decisions and post-drainage outcomes will require dedicated prospective study.
The current cross-view support count is not a calibrated uncertainty estimate and cannot establish that a generated connection is anatomically correct. Formal topology verification and repeated-sampling uncertainty analysis were not implemented; low-support or disconnected generated components therefore require explicit visualization and expert review.
Despite these limitations, the workflow provides a structured foundation for future research. The next phase should include multicenter data, standardized annotation of duct branches and strictures, external testing, confidence intervals, calibration analyses, ablation studies, reader studies with endoscopists, and comparison with conventional MRCP review for stent-length selection and drainage-route planning. Clinically meaningful endpoints should include agreement with expert procedural plans, adequacy of drainage, need for reintervention, adverse events, and time required for preprocedural planning. Concretely, the planned next steps are multicenter MRCP collection, testing across scanner vendors and acquisition protocols, external validation at independent institutions, and prospective clinical evaluation; consistent with contemporary guidance[24] for the development and reporting of clinical prediction models, these studies should report independent external validation, model calibration, and clinically meaningful endpoints rather than internal discrimination alone.
In conclusion, MRCP-based AI modeling of biliary strictures is feasible for quantitative 3D assessment and exploratory candidate stent-planning geometry. The present workflow can detect stricture presence internally, segment the biliary tree, restore simple interrupted ductal segments, and generate a centerline-guided tubular model. However, the method remains investigational. Robust reconstruction of complex hilar and intrahepatic anatomy, external validation, uncertainty-aware visualization, and prospective clinical testing are required before use in procedural decision-making or patient-specific device development.
| 1. | Elmunzer BJ, Maranki JL, Gómez V, Tavakkoli A, Sauer BG, Limketkai BN, Brennan EA, Attridge EM, Brigham TJ, Wang AY. ACG Clinical Guideline: Diagnosis and Management of Biliary Strictures. Am J Gastroenterol. 2023;118:405-426. [RCA] [PubMed] [DOI] [Full Text] [Cited by in Crossref: 114] [Cited by in RCA: 119] [Article Influence: 39.7] [Reference Citation Analysis (1)] |
| 2. | Facciorusso A, Crinò SF, Gkolfakis P, Spadaccini M, Arvanitakis M, Beyna T, Bronswijk M, Dhar J, Ellrichmann M, Gincul R, Hritz I, Kylänpää L, Martinez-Moreno B, Pezzullo M, Rimbaş M, Samanta J, van Wanrooij RLJ, Webster G, Triantafyllou K. Diagnostic work-up of bile duct strictures: European Society of Gastrointestinal Endoscopy (ESGE) Guideline. Endoscopy. 2025;57:166-185. [RCA] [PubMed] [DOI] [Full Text] [Cited by in Crossref: 38] [Cited by in RCA: 28] [Article Influence: 28.0] [Reference Citation Analysis (0)] |
| 3. | Rushbrook SM, Kendall TJ, Zen Y, Albazaz R, Manoharan P, Pereira SP, Sturgess R, Davidson BR, Malik HZ, Manas D, Heaton N, Prasad KR, Bridgewater J, Valle JW, Goody R, Hawkins M, Prentice W, Morement H, Walmsley M, Khan SA. British Society of Gastroenterology guidelines for the diagnosis and management of cholangiocarcinoma. Gut. 2023;73:16-46. [RCA] [PubMed] [DOI] [Full Text] [Full Text (PDF)] [Cited by in Crossref: 86] [Cited by in RCA: 71] [Article Influence: 23.7] [Reference Citation Analysis (3)] |
| 4. | Lee TH, Moon JH, Park SH. Endoscopic strategy and covered self-expandable metal stents for malignant hilar biliary obstruction. Clin Endosc. 2025. [RCA] [PubMed] [DOI] [Full Text] [Cited by in RCA: 1] [Reference Citation Analysis (0)] |
| 5. | Li G, Yang Y, Qu Y, Wang A, He Z, Peng Y. Safety and Effectiveness Analysis of Different Surgical Methods for Intrahepatic Bile Duct Stones. Ann Ital Chir. 2024;95:42-48. [PubMed] |
| 6. | Qumseya BJ, Jamil LH, Elmunzer BJ, Riaz A, Ceppa EP, Thosani NC, Buxbaum JL, Storm AC, Sawhney MS, Pawa S, Naveed M, Lee JK, Law JK, Kwon RS, Jue TL, Fujii-Lau LL, Fishman DS, Calderwood AH, Amateau SK, Al-Haddad M, Wani S. ASGE guideline on the role of endoscopy in the management of malignant hilar obstruction. Gastrointest Endosc. 2021;94:222-234.e22. [RCA] [PubMed] [DOI] [Full Text] [Cited by in Crossref: 59] [Cited by in RCA: 83] [Article Influence: 16.6] [Reference Citation Analysis (1)] |
| 7. | Mitsuhashi S, Aggarwal M, Chandrasekhara V. Palliative endoscopic biliary drainage for malignant hilar biliary obstruction. Clin Endosc. 2026. [RCA] [PubMed] [DOI] [Full Text] [Cited by in RCA: 1] [Reference Citation Analysis (0)] |
| 8. | Boodram RS, Vedantam S, Amin S, Bhalla S, Barkin JA, Mahfouz M, Kumar S. Outcomes associated with bilateral plastic versus metal stents for palliation in malignant hilar biliary obstruction: an updated meta-analysis from the United States. Clin Endosc. 2026;59:223-233. [RCA] [PubMed] [DOI] [Full Text] [Full Text (PDF)] [Cited by in RCA: 1] [Reference Citation Analysis (0)] |
| 9. | Vuppalanchi R, Are V, Telford A, Young L, Mouchti S, Ferreira C, Kettler C, Gromski M, Akisik F, Chalasani N. A composite score using quantitative magnetic resonance cholangiopancreatography predicts clinical outcomes in primary sclerosing cholangitis. JHEP Rep. 2023;5:100834. [RCA] [PubMed] [DOI] [Full Text] [Full Text (PDF)] [Cited by in Crossref: 22] [Cited by in RCA: 19] [Article Influence: 6.3] [Reference Citation Analysis (1)] |
| 10. | Robles-Medranda C, Verpalen I, Schulz D, Spadaccini M. Artificial Intelligence in Biliopancreatic Disorders: Applications in Cross-Sectional Imaging and Endoscopy. Gastroenterology. 2025;169:471-486. [RCA] [PubMed] [DOI] [Full Text] [Cited by in Crossref: 4] [Cited by in RCA: 7] [Article Influence: 7.0] [Reference Citation Analysis (0)] |
| 11. | Quarto G, Benassai G, Fernicola A. Using Artificial Intelligence to Enable Safer Laparoscopic Cholecystectomy: A Step Toward a Standardized Critical View of Safety and Reduced Bile Duct Injuries. Ann Ital Chir. 2025;96:1279-1281. [RCA] [PubMed] [DOI] [Full Text] [Cited by in RCA: 4] [Reference Citation Analysis (0)] |
| 12. | Sobek J, Medina Inojosa JR, Medina Inojosa BJ, Rassoulinejad-Mousavi SM, Conte GM, Lopez-Jimenez F, Erickson BJ. MedYOLO: A Medical Image Object Detection Framework. J Imaging Inform Med. 2024;37:3208-3216. [RCA] [PubMed] [DOI] [Full Text] [Full Text (PDF)] [Cited by in Crossref: 44] [Cited by in RCA: 10] [Article Influence: 5.0] [Reference Citation Analysis (0)] |
| 13. | Ogie P, Srikanthan P, Taiwo G, Ogundepo O. Hierarchical Fibroid Vision Transformer (HiFi-ViT): A Hierarchical Vision Transformer Framework for Accurate Uterine Fibroid Detection in Ultrasound Imaging. Cureus. 2026;18:e105197. [RCA] [PubMed] [DOI] [Full Text] [Full Text (PDF)] [Cited by in RCA: 1] [Reference Citation Analysis (0)] |
| 14. | Gomez JT, Rangnekar A, Williams H, Thompson HM, Garcia-Aguilar J, Smith JJ, Veeraraghavan H. Swin transformers are robust to distribution and concept drift in endoscopy-based longitudinal rectal cancer assessment. Proc SPIE Int Soc Opt Eng. 2025;13406:134061N. [RCA] [PubMed] [DOI] [Full Text] [Full Text (PDF)] [Cited by in Crossref: 3] [Cited by in RCA: 2] [Article Influence: 2.0] [Reference Citation Analysis (0)] |
| 15. | Sabanci K, Aslan B, Aslan MF. Medical Image Segmentation Methods: A Decision-Guided Survey Covering 2D/3D CNNs, Transformers, VLMs, SAM-Based Models and Diffusion Approaches. Bioengineering (Basel). 2026;13:555. [RCA] [PubMed] [DOI] [Full Text] [Full Text (PDF)] [Cited by in RCA: 3] [Reference Citation Analysis (0)] |
| 16. | Wang Y, Fu H, Cao Q, Wang S, Chen Z, Liang F. S2LIC: Learned image compression with the SwinV2 block, Adaptive Channel-wise and Global-inter attention Context. Neural Netw. 2025;189:107590. [RCA] [PubMed] [DOI] [Full Text] [Cited by in RCA: 1] [Reference Citation Analysis (0)] |
| 17. | Becq A, Szewczyk J, Salin G, Chartier M, Chaput U, Leenhardt R, Dray X, Arrive L, Camus M. ERCP 2.0: Biliary 3D-reconstruction in patients with malignant hilar stricture. Clin Res Hepatol Gastroenterol. 2023;47:102172. [RCA] [PubMed] [DOI] [Full Text] [Cited by in RCA: 7] [Reference Citation Analysis (0)] |
| 18. | Yeung M, Sala E, Schönlieb CB, Rundo L. Unified Focal loss: Generalising Dice and cross entropy-based losses to handle class imbalanced medical image segmentation. Comput Med Imaging Graph. 2022;95:102026. [RCA] [PubMed] [DOI] [Full Text] [Full Text (PDF)] [Cited by in Crossref: 36] [Cited by in RCA: 184] [Article Influence: 36.8] [Reference Citation Analysis (0)] |
| 19. | Rema P, Moorthy S, K P S, Bright B, B R M. An attention-guided residual 3D U-net with focal Tversky-Dice loss for multi-modal pancreatic tumor segmentation using synthetic volumetric imaging. Cancer Treat Res Commun. 2026;47:101146. [RCA] [PubMed] [DOI] [Full Text] [Cited by in RCA: 1] [Reference Citation Analysis (0)] |
| 20. | Allan A, Kealley C, Squelch A, Wong YH, Yeong CH, Sun Z. Patient-specific 3D printed model of biliary ducts with congenital cyst. Quant Imaging Med Surg. 2019;9:86-93. [RCA] [PubMed] [DOI] [Full Text] [Cited by in Crossref: 10] [Cited by in RCA: 19] [Article Influence: 2.7] [Reference Citation Analysis (0)] |
| 21. | Boyer CJ, Boktor M, Samant H, White LA, Wang Y, Ballard DH, Huebert RC, Woerner JE, Ghali GE, Alexander JS. 3D Printing for Bio-Synthetic Biliary Stents. Bioengineering (Basel). 2019;6:16. [RCA] [PubMed] [DOI] [Full Text] [Full Text (PDF)] [Cited by in Crossref: 39] [Cited by in RCA: 34] [Article Influence: 4.9] [Reference Citation Analysis (2)] |
| 22. | Thomas J, Patel S, Troop L, Guru R, Faist N, Bellott BJ, Esterlen BA. 3D Printed Model of Extrahepatic Biliary Ducts for Biliary Stent Testing. Materials (Basel). 2020;13:4788. [RCA] [PubMed] [DOI] [Full Text] [Full Text (PDF)] [Cited by in Crossref: 5] [Cited by in RCA: 10] [Article Influence: 1.7] [Reference Citation Analysis (0)] |
| 23. | Gildea TR, Young BP, Machuzak MS. Application of 3D Printing for Patient-Specific Silicone Stents: 1-Year Follow-Up on 2 Patients. Respiration. 2018;96:488-494. [RCA] [PubMed] [DOI] [Full Text] [Cited by in Crossref: 50] [Cited by in RCA: 43] [Article Influence: 5.4] [Reference Citation Analysis (0)] |
| 24. | Li J, Chen Z, Wang Q, Du X, Zhang J. Radiomics Model Based on Preoperative Contrast-enhanced CT and Clinical Features for Predicting Early Recurrence of Hepatocellular Carcinoma. Ann Ital Chir. 2025;96:771-782. [RCA] [PubMed] [DOI] [Full Text] [Cited by in Crossref: 2] [Cited by in RCA: 4] [Article Influence: 4.0] [Reference Citation Analysis (0)] |