Preoperative MVI prediction in intrahepatic cholangiocarcinoma via deep learning analysis of intratumoral and peritumoral features on multi-sequence MRI

Abstract

Background

To explore the prediction effectiveness of intratumoral and peritumoral regions in multi-sequence magnetic resonance imaging (MRI) based on deep learning（DL）, determine the optimal region of interest (ROI), and develop an efficient preoperative prediction model for microvascular invasion (MVI) in intrahepatic cholangiocarcinoma (ICC).

Methods

Clinical, pathological, and preoperative MRI data of 200 patients with ICC confirmed by surgical pathology were retrospectively collected. All patients had preoperative contrast-enhanced MRI scans. Patients were randomized in a 7:3 ratio into a training cohort (n = 140) and a validation cohort (n = 60). Univariate and multivariate logistic regression were used to screen for independent clinical imaging risk factors for MVI. Deep learning features were extracted from three different ROIs (ROI_tumor, ROI_peri10mm, and ROI_{tumor+peri10mm}) of the maximum transverse cross-section of tumors in six MRI sequence images using a pre-trained ResNet-18. Construct clinical-imaging model, tumor model, peri10mm model, tumor₊peri10mm model, and combined model with deep learning score and clinical-imaging features based on logistic regression (LR) classifiers for predicting MVI. Using the receiver operating characteristic (ROC) curve and calculating the area under the curve (AUC), the models’ effectiveness in predicting MVI was evaluated. The clinical utility of these models were evaluated using decision curve analysis (DCA). Interpretability analysis of the best predictive model using the Shapley additive explanation(SHAP) method.

Results

Tumor size, intrahepatic duct dilatation, and arterial edge enhancement ratio were independent MVI predictors. Superior prediction performance was shown by the tumor₊peri10mm model (training set AUC 0.932, validation set AUC 0.872). The model’s performance was further enhanced by the incorporation of clinical imaging features (AUC 0.883, sensitivity 84.2%, specificity 78.0%). And all of the differences with the other prediction models showed statistical significance, based on the DeLong test (all p＜0.05). SHAP-based explainability research highlighted the significance of deep learning score.

Conclusions

In summary, the combined model integrating clinical imaging features with deep learning features derived from multi-sequence MRI within the tumor and at a 10-millimeter depth around the tumor demonstrated favorable results in predicting the MVI status of ICC patients. This approach may aid in guiding relevant clinical decisions and improving patient prognosis.

Clinical trial number

Not applicable.

Supplementary information

The online version contains supplementary material available at 10.1186/s12880-025-02107-z.

Introduction

Primary malignant liver tumors include hepatocellular carcinoma (HCC), intrahepatic cholangiocarcinoma (ICC), and mixed hepatocellular-cholangiocarcinoma [1]. Among these, ICC ranks as the second most common type, accounting for approximately 10%–15% of primary liver cancers, with a rising incidence in recent years [2]. Compared to HCC, treatment options for ICC are more limited and challenging. Despite ongoing developments in multidisciplinary comprehensive therapies, radical hepatectomy (R0 resection) remains the sole curative treatment, particularly for patients with early-stage, resectable, localized lesions. However, owing to ICC’s invasive biological behavior, frequent detection at advanced stages, and the absence of effective systemic therapeutic agents, most patients are ineligible for surgical intervention, and only 20%–30% of patients undergo surgery. Even after surgical resection, the five-year postoperative recurrence rate remains high (50–70%) [3, 4].

Microvascular invasion (MVI) refers to the histological finding of tumor cell nests within endothelial-lined vessels, predominantly observed in small portal venous branches within the peritumoral hepatic tissues [5]. The presence of MVI significantly correlates with poorer prognosis, contributing to increased recurrence rates and decreased survival in patients with ICC [6, 7]. Chang et al. [7] demonstrated that ICC patients exhibiting MVI experienced significantly reduced overall survival (OS). Accurate preoperative prediction of MVI can enable clinicians to develop personalized therapeutic strategies, such as expanding surgical resection margins, performing anatomical resections, or administering neoadjuvant therapies to MVI-positive individuals, thereby enhancing survival rates and reducing metastatic recurrence risks [8–11]. Currently, however, postoperative pathological diagnosis remains the gold standard for MVI identification, while preoperative biopsy is invasive and limited by potential false-positive outcomes due to sampling variability. Previous research identified ADC values [12], tumor size [13–16], intrahepatic bile duct dilation [13, 16, 17], hepatic vascular invasion signs [16], and serological markers [14–16] as factors associated with MVI through multivariate analyses of clinical and imaging data. Nonetheless, conventional imaging assessments are subjective, restricting their clinical applicability, and inexperienced radiologists may overlook subtle imaging features relevant to ICC. Furthermore, predictive models relying solely on these features have demonstrated inadequate diagnostic performance, achieving areas under the curve (AUCs) ranging from 0.522 to 0.782 [12, 14, 15].

Radiomics enables the high-throughput extraction and analysis of numerous imaging features related to tumor biology from medical images, reflecting microscopic lesion changes undetectable by the naked eye, thus revealing tumor heterogeneity. In recent years, radiomics has demonstrated utility in the differential diagnosis, therapeutic efficacy assessment, and prognostic prediction of liver tumors [18–21]. However, traditional radiomics continues to face several challenges. First, accurate image segmentation requires manual delineation by experienced radiologists, which is labor-intensive, time-consuming, and susceptible to inter-observer variability. Second, it relies on the calculation of pre-designed imaging features such as wavelet, shape, and texture. With continuous advancements in artificial intelligence (AI), deep learning (DL) is becoming increasingly popular in medical imaging, as it allows powerful quantitative analysis of complex image features. DL automatically evaluates medical images without predefined radiological feature extraction and may even identify previously unknown relevant features, thereby demonstrating substantial potential for efficient diagnosis and prediction [22]. Zhang et al. [23] validated across multiple institutions that DL shows greater robustness than traditional radiomics in preoperatively predicting MVI in HCC. Recent studies have indicated that applying DL to ICC has provided promising outcomes for diagnosis and prognosis [22, 24, 25].

Currently, most DL-based ICC studies focus primarily on the lesion itself, and few employ peritumoral DL features to predict MVI in ICC [26]. Given that MVI predominantly occurs in liver tissue surrounding the tumor, the peritumoral region might offer critical predictive information [5]. Previous studies have demonstrated that imaging characteristics of the tumor microenvironment provide essential insights into the biological properties of malignancies [17, 21]. However, intratumoral and peritumoral regions typically lack clear contrast. To explore the impact of DL features derived from tumor and peritumoral regions on MVI prediction, this study aims to extract DL features from various regions (intratumoral alone, peritumoral alone, and combined intratumoral-peritumoral regions) and determine the optimal region of interest (ROI) for preoperative MVI status prediction.

Materials and methods

Selection of patients

This retrospective analysis obtained ethical approval from the ethics committees of Xuzhou Central Hospital (XZXY-LK-20241105-0173) and Zhongshan Hospital of Fudan University (B2021-325 R), and the requirement for informed consent was waived. Given the retrospective design, patient identities were fully anonymized prior to conducting the analyses. We retrospectively collected clinical, pathological, and MRI data from 382 patients who underwent surgical resection and received pathologically confirmed diagnoses of intrahepatic cholangiocarcinoma (ICC) from January 2016 to October 2022. Inclusion criteria included: (1) radical surgical resection with ICC diagnosis confirmed histologically; (2) availability of comprehensive pathological data; and (3) tumor diameter ≥10 mm. Exclusion criteria encompassed: (1) prior history of liver tumor therapy; (2) MRI examination performed more than 30 days prior to surgery; (3) presence of multiple (≥2) hepatic lesions; (4) inadequate MRI image quality due to respiratory artifacts. After applying these exclusion criteria, the final study population comprised 200 patients. These patients were subsequently divided into training (n = 140) and validation (n = 60) groups using a randomized allocation ratio of 7:3 (Fig. 1).

Fig. 1.

Flowchart for patient inclusion and exclusion

Clinical and pathological information

Relevant clinical data, including age, sex, hepatitis B history, and serum biomarkers (alpha-fetoprotein [AFP], carcinoembryonic antigen [CEA], cancer antigen 19–9 [CA19-9]), were retrieved from the electronic medical record systems. All tumor specimens were collected from hepatectomy resections according to a standardized seven-point baseline sampling protocol [27]. MVI status was independently assessed by two pathologists at each institution, with disagreements resolved through discussion.

Acquisition of MRI images

All patients underwent multi-sequence abdominal MRI, including plain scans and three-phase contrast-enhanced imaging. To minimize respiratory motion artifacts, MRI scans were conducted after breath-hold training. MRI examinations were performed using various scanners, including: a Philips 1.5T Achieva (Philips Medical Systems, Best, Netherlands); two United Imaging scanners—1.5T uMR560 and 3.0T uMR770 (United Imaging Healthcare, Shanghai, China); Siemens scanners including 1.5T Magnetom Avanto, 1.5T Magnetom Aera, and 3.0T Magnetom Verio (Siemens Healthcare, Erlangen, Germany); and three GE 3.0T Discovery 750 units (GE Healthcare, Milwaukee, WI, USA). The MRI imaging protocols consisted of T1-weighted imaging (T1WI), fat-suppressed T2-weighted imaging (T2WI-FS), diffusion-weighted imaging (DWI; b = 0, 800 mm²/s), and dynamic contrast-enhanced T1-weighted imaging (CE-T1WI). During contrast-enhanced imaging, intravenous administration of Gd-DTPA (0.2 mmol/kg) was performed, followed by a saline flush (20 ml) delivered at 2 ml/s using a power injector. Image acquisition phases included arterial (20–30 s), portal venous (60–70 s), and delayed (180 s) scans. Comprehensive MRI acquisition parameters are detailed in Table S1.

Features of MRI imaging

Two experienced abdominal radiologists independently evaluated the images through a double-blind method, and disagreements were resolved by a third radiologist with over 20 years of experience. The imaging features assessed included: (a) tumor size (maximum axial diameter); (b) tumor morphology (round/hemispherical/ovoid, lobulated, irregular); (c) intrahepatic bile duct dilatation; (d) visible vascular penetration; (e) arterial edge enhancement ratio (ratio of maximal border enhancement thickness to the maximum diameter in the arterial phase); (f) arterial-phase enhancement patterns (peripheral rim hyperenhancement, diffuse homogeneous hyperenhancement, heterogeneous mosaic enhancement, minimal or absent enhancement); (g) overall enhancement patterns (rapid wash-in and wash-out, progressive enhancement, or persistent enhancement); (h) hepatic capsular retraction; (i) peripheral hepatic enhancement; (j) relative tumor signal intensity (SI) on T1WI and T2WI-FS compared to adjacent liver tissue (categorized as low, equal, or high); (k) presence of a target sign on T2WI-FS and DWI sequences (characterized by peripheral hyperintensity surrounding central iso- or hypointensity); and (l) LI-RADS classification based on LI-RADSv2018 [28].

Analysis for DL

Workflow

DL analysis consisted of three sequential steps (Fig. 2): (1) segmentation and expansion of tumor regions; (2) extraction and selection of relevant imaging features; and (3) construction and validation of predictive models.

Fig. 2.

Study flowchart of the deep learning analysis

Tumor segmentation

All original MRI images underwent preprocessing to minimize center effects caused by variations in institutions and scanners. T2-weighted imaging (T2WI), which provides clear anatomical structures, was used as the reference sequence for alignment. A rigid registration method based on mutual information was applied to achieve cross-sequence spatial alignment. All MRI sequences underwent resampling to a uniform spatial resolution of 1 mm × 1 mm × 1 mm to ensure precise anatomical alignment across modalities (achieving mean registration errors below 1 mm). Subsequently, bias field correction was performed on all images. Signal intensity normalization to the [0,1] interval was applied using min–max normalization based on intensity ranges within tumor parenchymal regions, deliberately excluding necrotic and artifact-containing areas to maintain signal uniformity across scans. Two experienced radiologists independently delineated the region of interest (ROI) manually on images showing the largest transverse diameter of each tumor across six MRI sequences, including T2WI, DWI, T1WI, and the three phases of CE-T1WI (arterial phase [T1WI-AP], portal venous phase [T1WI-PVP], and delayed phase [T1WI-DP]), using the 3D Slicer software platform (version 4.10.2). Both radiologists remained blinded to histopathological results during segmentation. The resulting tumor ROI was expanded using the BinaryDilateImageFilter function in the SimpleITK library (Python) to extend the peritumoral boundary by 10 mm. Regions beyond the hepatic parenchyma were manually removed. Three types of ROIs were generated: the tumor ROI (ROI_tumor), the 10-mm peritumoral ROI (ROI_peri10mm), and the combined tumor plus peritumoral ROI (ROI_{tumor+peri10mm}). In cases of disagreement during ROI delineation, the two radiologists reached consensus through discussion. The segmentation consistency between observers was assessed using the intersection-over-union (IoU) metric for both tumor and marginal regions. All IoU values exceeded 0.85, indicating high inter-observer agreement and effectively minimizing subjective variability, thereby providing standardized data for subsequent MVI prediction model development.

Feature extraction and selection

The tumor’s maximum transverse cross-sectional image was selected as the input for the DL model. This study employed the pre-trained model torchvision.models.resnet18 within the PyTorch framework to extract DL features from three distinct regions of interest (ROI_tumor, ROI_peri10mm, and ROI_{tumor+peri10mm}). Subsequently, features extracted from six MRI sequences corresponding to the same ROI were fused, followed by normalization using Z-score transformation. Feature selection was conducted using the sklearn library to remove redundant features and optimize the feature set. Features with variance greater than 0.8 were initially selected applying the method of variance threshold; features with p < 0.05 were subsequently selected using the SelectKBest; Lastly, the 10-fold cross-validated least absolute shrinkage and selection operator (LASSO) was used to further select the most representative features.LASSO determined the parameter alpha according to the size of the mean square error (MSE) value, retained features with non-zero coefficients, and determined the features and their coefficients (Fig.S1).LASSO feature selection and model training employ independent cross-validation mechanisms. The former determines feature subsets, whilst the latter optimizes classifier parameters. No data or parameters are shared between the two.The patient DL score was produced by linear weighting the final screened DL features based on their respective feature regression coefficients. The scoring formula was as follows:

Model construction and evaluation

Initial variable screening employed univariate analysis, followed by multivariate logistic regression. Final variable retention comprehensively considered statistical association, biological plausibility, clinical relevance.Subsequently, these predictors informed the construction of the clinical-imaging prediction model, which was established using logistic regression (LR). Following feature extraction and selection, features from six MRI sequences within each ROI were combined and independently modeled, yielding three separate predictive models: tumor, peri10mm, and combined tumor+peri10mm models. Additionally, an integrated clinical-imaging model was developed by combining DL scores derived from tumor and peri10mm regions with relevant clinical variables. Hyperparameter optimization was conducted through GridSearchCV from the scikit-learn package, employing an automated parameter search using 5-fold cross-validation (CV = 5). After determining the best hyperparameters, the model was retrained using the entire training cohort. The final model performance with optimal parameters was subsequently evaluated within the validation cohort. Predictive capabilities of each model for ICC-MVI were quantified by calculating sensitivity, specificity, accuracy, and AUC. Clinical utility and applicability were assessed using decision curve analysis (DCA) and calibration curves. Model interpretability was further enhanced using Shapley Additive Explanations (SHAP), quantifying individual feature contributions, identifying critical predictors, and clarifying clinical relevance. SHAP plots were produced to visualize feature impact on predictions. Additionally, Gradient-weighted Class Activation Mapping (Grad-CAM) was applied to visually depict regions contributing to the DL model’s predictions.

Analysis of statistical

Statistical analyses were conducted using R software (version 4.2.1) and Python (version 3.6.8). Categorical variables were evaluated by Chi-square or Fisher’s exact tests. Continuous variable normality was determined using Kolmogorov-Smirnov testing, normally distributed parameters (mean±SD) underwent Student’s t-test, the Mann-Whitney U-test was used for data that didn’t fit the normal distribution, and skewed distributions data were expressed as medians (upper and lower quartiles) [M (P25, P75)]. Model discrimination was compared via DeLong’s method for AUC differences. Statistical significance threshold was set at α = 0.05 (two-tailed).

Results

Baseline features

Table 1 summarizes baseline clinical-imaging characteristics related to MVI presence in the training and validation cohorts. A total of 200 ICC patients were enrolled, including 140 patients (44 MVI-positive, 96 MVI-negative) in the training group and 60 patients (19 MVI-positive, 41 MVI-negative) in the validation group. No significant differences in MVI prevalence were observed between cohorts (p > 0.05). Univariate analyses demonstrated significant correlations between ICC-MVI and serum CEA levels (p = 0.041), tumor size (p = 0.001), intrahepatic bile duct dilatation (p = 0.001), and arterial edge enhancement ratio (p = 0.001). Multivariate logistic regression analysis identified intrahepatic bile duct dilatation (p = 0.004; OR = 3.46, 95% CI = 1.48–8.09), arterial edge enhancement ratio (p = 0.033; OR = 0.03, 95% CI = 0–0.76), and tumor size (p = 0.05; OR = 1.02, 95% CI = 1–1.03) as independent predictive factors for MVI (Table 2). Finally, the clinical-imaging prediction model based on logistic regression yielded AUC values of 0.748 (95% CI: 0.664–0.832) in the training cohort and 0.734 (95% CI: 0.587–0.882) in the validation cohort (Table 3).

Table 1.

Baseline features

Characteristics	Training cohort (n = 140)			Validation cohort (n = 60)			p-Inter
	MVI (-), (n = 96)	MVI (+), (n = 44)	p-Intra	MVI (-), (n = 41)	MVI (+), (n = 19)	p-Intra
Clinical features
Gender			1.00			0.49	0.41
Female	25 (26%)	12 (27.3%)		12 (29.3%)	8 (42.1%)
Male	71 (74%)	32 (72.7%)		29 (70.7%)	11 (57.9%)
Age, years (mean ± SD)	59.2 ± 11.1	61.9 ± 10.9	0.18	63.6 ± 11.5	60.9 ± 11.1	0.40	0.12
HBV			0.27			0.07	0.71
Negative	50 (52.1%)	28 (63.6%)		25 (61%)	6 (31.6%)
Positive	46 (47.9%)	16 (36.4%)		16 (39%)	13 (68.4%)
AFP			0.96			0.03	0.29
＜20 ng/ml	81 (84.4%)	38 (86.4%)		40 (97.6%)	15 (78.9%)
≥20 ng/ml	15 (15.6%)	6 (13.6%)		1 (2.4%)	4 (21.1%)
CEA			0.06			0.45	0.60
＜5 ng/ml	82 (85.4%)	31 (70.5%)		36 (87.8%)	15 (78.9%)
≥5 ng/ml	14 (14.6%)	13 (29.5%)		5 (12.2%)	4 (21.1%)
CA199			0.28			1.00	1.00
＜34 U/ml	59 (61.5%)	22 (50%)		23 (56.1%)	11 (57.9%)
≥34 U/ml	37 (38.5%)	22 (50%)		18 (43.9%)	8 (42.1%)
MR imaging features
Tumor size, mm*	39.1 (22.8, 55.5%)	54.2 (36.8, 73.6%)	< 0.001	40.7 (30.7, 48.7%)	49.8 (39.5, 82.4%)	0.02	0.78
Tumor morphology			0.06			0.48	0.95
(Hemi-)spherical and oval	41 (42.7%)	11 (25%)		15 (36.6%)	7 (36.8%)
Lobulated	40 (41.7%)	20 (45.5%)		20 (48.8%)	7 (36.8%)
Irregular	15 (15.6%)	13 (29.5%)		6 (14.6%)	5 (26.3%)
SI on T1WI			0.25			1.00	0.52
Low	94 (97.9%)	41 (93.2%)		41 (100%)	19 (100%)
Moderate	2 (2.1%)	2 (4.5%)		0 (0%)	0 (0%)
High	0 (0%)	1 (2.3%)		0 (0%)	0 (0%)
SI on T2WI-FS			1.00			1.00	0.30
Low				1 (2.4%)	0 (0%)
Moderate	2 (2.1%)	1 (2.3%)		2 (4.9%)	0 (0%)
High	94 (97.9%)	43 (97.7%)		38 (92.7%)	19 (100%)
Target sign on T2WI-FS			0.71			1.00	0.30
Negative	63 (65.6%)	31 (70.5%)		24 (58.5%)	11 (57.9%)
Positive	33 (34.4%)	13 (29.5%)		17 (41.5%)	8 (42.1%)
Target sign on DWI			0.60			0.84	0.14
Negative	55 (57.3%)	28 (63.6%)		20 (48.8%)	8 (42.1%)
Positive	41 (42.7%)	16 (36.4%)		21 (51.2%)	11 (57.9%)
LI-RADS			0.46			0.55	0.74
LR-5	8 (8.3%)	5 (11.4%)		2 (4.9%)	2 (10.5%)
LR-M	84 (87.5%)	38 (86.4%)		36 (87.8%)	17 (89.5%)
LR-TIV	0 (0%)	1 (2.3%)		0 (0%)	0 (0%)
LR-3	1 (1%)	0 (0%)		0 (0%)	0 (0%)
LR-4	3 (3.1%)	0 (0%)		3 (7.3%)	0 (0%)
Intrahepatic duct dilatation			< 0.001			0.07	0.39
Negative	71 (74%)	19 (43.2%)		27 (65.9%)	7 (36.8%)
Positive	25 (26%)	25 (56.8%)		14 (34.1%)	12 (63.2%)
Hepatic capsular retraction			0.50			0.86	0.13
Negative	64 (66.7%)	26 (59.1%)		22 (53.7%)	9 (47.4%)
Positive	32 (33.3%)	18 (40.9%)		19 (46.3%)	10 (52.6%)
Arterial edge enhancement ratio (mean (SD))	0.2 (0.1, 0.4)	0.1 (0, 0.2)	< 0.001	0.1 (0.1, 0.2)	0.1 (0, 0.2)	0.48	0.04
Arterial phase enhancement pattern			0.02			0.68	0.43
1	11 (11.5%)	2 (4.5%)		4 (9.8%)	0 (0%)
2	80 (83.3%)	34 (77.3%)		31 (75.6%)	16 (84.2%)
3	5 (5.2%)	5 (11.4%)		3 (7.3%)	2 (10.5%)
4	0 (0%)	3 (6.8%)		3 (7.3%)	1 (5.3%)
Enhancement pattern			0.78			0.10	0.04
Gradual and filling	13 (13.5%)	7 (15.9%)		3 (7.3%)	2 (10.5%)
Arterial and persistent	73 (76%)	34 (77.3%)		26 (63.4%)	16 (84.2%)
Wash-in and wash-out	10 (10.4%)	3 (6.8%)		12 (29.3%)	1 (5.3%)
Visible vessel penetration			0.48			0.76	0.26
Negative	47 (49%)	18 (40.9%)		14 (34.1%)	8 (42.1%)
Positive	49 (51%)	26 (59.1%)		27 (65.9%)	11 (57.9%)
Peripherally hepatic enhancement			0.63			0.03	0.62
Negative	45 (46.9%)	18 (40.9%)		12 (29.3%)	12 (63.2%)
Positive	51 (53.1%)	26 (59.1%)		29 (70.7%)	7 (36.8%)

Table 2.

Univariate and multivariate analyses of predicted factors associated with MVI status in ICC

Characteristics	Univariate			Multivariate
	OR	95% CI	p-value	OR	95% CI	p-value
Gender	0.94	0.42–2.10	0.878
Age	1.02	0.99–1.06	0.182
HBV	0.62	0.30–1.29	0.203
AFP	0.85	0.31–2.37	0.760
CEA	2.46	1.04–5.81	0.041	1.7	0.62–4.63	0.300
CA199	1.59	0.78–3.28	0.204
Tumor size	1.03	1.01–1.04	0.001	1.02	1–1.03	0.05
Tumor morphology	1.86	0.79–4.38	0.154
SI on T1WI	2.29	0.31–16.84	0.415
SI on T2WI-FS	0.91	0.08–10.37	0.943
Target sign on T2WI-FS	0.80	0.37–1.73	0.573
Target sign on DWI	0.77	0.37–1.60	0.479
LI-RADS	0.72	0.22–2.36	0.592
Intrahepatic duct dilatation	3.74	1.76–7.92	0.001	3.46	1.48–8.09	0.004
Hepatic capsular retraction	1.38	0.66–2.89	0.386
Arterial edge enhancement ratio	0.01	0.00–0.11	0.001	0.03	0–0.76	0.033
Arterial phase enhancement pattern	2.34	0.49–11.11	0.286
Enhancement pattern	0.86	0.32–2.36	0.777
Visible vessel penetration	1.39	0.67–2.85	0.376
Peripherally hepatic enhancement	1.27	0.62–2.62	0.510

Table 3.

Each model’s training and validation cohort results using LR modeling

Model	Classifier	AUC	95%CI	acc	sen	spec	precision	F1-score
Training cohort	clinical-imaging model	0.748	0.664–0.832	0.65	0.659	0.646	0.46	0.542
	tumor model	0.878	0.819–0.936	0.793	0.795	0.792	0.636	0.707
	peri10mm model	0.909	0.860–0.959	0.814	0.864	0.792	0.655	0.745
	tumor+peri10mm model	0.932	0.889–0.974	0.85	0.841	0.854	0.725	0.779
	combined model	0.880	0.825–0.936	0.814	0.841	0.802	0.661	0.74
Validation cohort	clinical-imaging model	0.734	0.587–0.882	0.600	0.684	0.561	0.419	0.52
	tumor model	0.807	0.688–0.927	0.717	0.789	0.683	0.536	0.638
	peri10mm model	0.786	0.633–0.938	0.717	0.684	0.732	0.542	0.605
	tumor+peri10mm model	0.872	0.774–0.970	0.767	0.842	0.732	0.593	0.696
	combined model	0.883	0.793–0.974	0.800	0.842	0.780	0.640	0.727

DL feature selection

For each patient, 512 DL features were extracted using the average pooling layer of ResNet-18 from three distinct ROIs (ROI_tumor, ROI_peri10mm, and ROI_{tumor+peri10mm}) across six MRI sequences. Features from the six sequences within the same ROI were fused, resulting in 3072 features (6 × 512) per region. Initially, the variance threshold method retained 1378 features from each region; subsequent univariate selection reduced these to 755 (ROI_tumor), 836 (ROI_peri10mm), and 738 (ROI_{tumor+peri10mm}) features. Finally, LASSO analysis selected 31, 29, and 32 features, respectively, that were strongly correlated with MVI (Fig. S1, S2, and Table S2).

Model construction and evaluation

Using logistic regression, four distinct models were established: the tumor model, peri10mm model, tumor+peri10mm model, and the combined model (Table 3, Fig. 4A). Across both the training and validation cohorts, all deep learning (DL) models demonstrated markedly higher predictive capability for identifying ICC-MVI compared with the clinical-imaging model (AUC range: 0.878–0.932 vs. 0.748 in training; 0.807–0.872 vs. 0.734 in validation). Among these, the tumor+peri10mm model achieved the best overall performance, with AUC values of 0.932 and 0.872 in the training and validation cohorts, respectively, surpassing the tumor-only (AUC = 0.878, 0.807) and peri10mm-only (AUC = 0.909, 0.872) models. In the validation set, the integrated combined model, incorporating clinical-imaging predictors with DL-derived scores, further enhanced discriminative accuracy (AUC = 0.883, 95% CI: 0.793–0.974), yielding a sensitivity of 84.2% and a specificity of 78.0%. The DL score differed significantly between MVI-positive and MVI-negative groups (p < 0.05, Fig. S3). Comparative analysis using DeLong’s test confirmed statistically significant differences in discriminative performance between the combined model and other models (all p < 0.05, Fig. 3). Calibration plots indicated strong agreement between predicted probabilities and actual outcomes in both datasets. Notably, the combined and tumor+peri10mm models exhibited the lowest Brier scores (0.163), reflecting the highest calibration accuracy in the validation cohort (Fig. 4B). Furthermore, decision curve analysis (DCA) demonstrated that the combined model provided the greatest net clinical benefit within risk thresholds ranging from 45% to 60% in the validation dataset (Fig. 4C).

Fig. 4.

Plot of results of model evaluation. (A) ROC curve (B) Calibration curve. (C) Decision curve

Fig. 3.

Delong test results for the two-by-two model (A) training cohort, (B) validation cohort

Visualization of DL features

To establish a reliable basis for clinical interpretation, heatmaps were generated to visualize the DL features extracted from six MRI sequences using the pre-trained ResNet-18 model. Warm colors (red/orange) represent regions of high attention, while cooler colors (blue/purple) indicate areas of low attention. The heatmap revealed that the six sequences displayed varying degrees of high-attention regions in both intratumoral and peritumoral areas (Fig. 5).

Fig. 5.

The pre-trained ResNet-18 model was used to extract features in the tumor₊peri10mm region for six sequential images. A representative example of an attention heat map was generated using the gradient-weighted class activation mapping (Grad-CAM) method

Shap analysis

This study utilized SHAP analysis to quantify the contribution of each feature to the model’s predictions and to identify critical predictive features. The SHAP bar plot (Fig. 6A) ranked the four most influential features (DL score, arterial edge enhancement ratio, intrahepatic bile duct dilatation, and tumor size) according to their weights. The DL score had the highest weight, thus exerting the greatest impact on model output. The SHAP summary plot (Fig. 6B) depicted each feature’s positive (red) or negative (blue) influence on the predicted probability. Additionally, two representative cases illustrate model interpretability: one patient was MVI-positive, with a predicted probability of 0.724 (Fig. 6C), and another was MVI-negative, with a predicted probability of 0.808 (Fig. 6D).

Fig. 6.

Shap analysis graph. (A) The shap bar plot, with the horizontal axis being the average shap value, indicates how important the feature is to the model’s predictions. (B) The shap summary plot demonstrates the distribution of the impact of each feature on the model’s predictions, with the color indicating the high or low value of the feature and the x-axis indicating the size of the shap value. (C) This is a patient who is MVI-positive, and the model predicts a probability of MVI-positive in this sample of 0.724. (D) This is a patient with a negative MVI and the model predicts a probability of 0.808 for a negative MVI

Discussion

This study extracted DL features from multiple regions in multi-sequence MRI images and combined them with clinical-imaging characteristics to develop a composite predictive model, aiming to identify the optimal model for preoperative prediction of ICC-MVI. First, the study identified tumor size, arterial edge enhancement ratio, and intrahepatic bile duct dilatation as important predictors of MVI, subsequently constructing a clinical-imaging model. Second, the tumor+peri10mm model outperformed the tumor-only, peri10mm-only, and clinical-imaging models, achieving AUCs of 0.932 and 0.872 in the training and validation cohorts, respectively. Moreover, the integrated combined model, incorporating both DL features and clinical-imaging parameters, further improved predictive reliability (AUC = 0.883, 95% CI: 0.793–0.974; sensitivity = 84.2%; specificity = 78.0%) in the validation set. To enhance clinical interpretability, SHAP value analysis, calibration curves, and DCA were applied, providing quantitative insight into feature contributions and overall model robustness.

Previous studies indicated that MVI is a significant factor contributing to poor postoperative survival in ICC and serves as a reference marker for postoperative adjuvant chemotherapy [6, 7, 11]. The peritumoral region is particularly vulnerable to MVI invasion, as its vascular structures constitute the primary route for portal tumor thrombosis and metastasis [29]. Emerging evidence suggests that analyzing the peritumoral region can provide crucial insights into tumor biology, enhance prognostic accuracy, and inform clinical management [30–32]. Therefore, evaluating the predictive value of peritumoral tissue for MVI is necessary. MVI predominantly occurs within the 10-mm peritumoral zone [33]. Zhou et al. reported that 79.9% of MVI occurrences were located within 10 mm of the tumor margin, suggesting that this margin distance may be clinically optimal [8]. Liu et al. [34] developed a CT radiomics model incorporating a 10-mm peritumoral region to predict ICC-MVI. Similarly, Chen et al. found that imaging features extracted from the tumor and its surrounding 10-mm region offered superior predictive accuracy for HCC treatment response to initial transarterial chemoembolization therapy [35]. Although previous studies have recognized that the 10-mm peritumoral region effectively predicts liver tumors’ characteristics and treatment response, and it represents the optimal range for capturing information related to tumor invasion [17, 29, 32, 34, 35], most studies evaluated peritumoral regions only in combination with intratumoral regions, thus may obscure the independent predictive significance of the peritumoral microenvironment. To overcome this limitation, the present study independently assessed the 10-mm peritumoral region, validating whether it possesses distinct predictive efficacy compared to tumor tissue alone. This independent assessment was motivated by two main considerations: (1) MVI predominantly occurs within the 10-mm peritumoral zone [33]. Notably, features predictive of MVI may appear earlier in peritumoral tissue than within tumor parenchyma. (2) Independently evaluating the 10-mm peritumoral region helps clarify whether this region offers distinct predictive value or redundancy relative to intratumoral tissue. The current findings revealed that the predictive performance of the 10-mm peritumoral region alone (AUC = 0.786) was lower than the intratumoral region (AUC = 0.807) yet substantially exceeded random prediction(AUC = 0.5), confirming its independent predictive significance. These peritumoral features may reflect early microenvironmental changes associated with tumor invasion, offering clinicians novel biomarkers to identify high-risk patients at an early stage. Furthermore, this study demonstrated that integrating intratumoral and peritumoral information (tumor+peri10mm model) provided superior predictive accuracy (AUC = 0.932, training; 0.872, validation) compared to either region alone, consistent with previous literature [5, 36]. Combining complementary intratumoral and peritumoral information reduces prediction bias resulting from incomplete data inherent in single-region analyses. For instance, even if certain intratumoral features suggest low invasiveness, vascular invasion could already be evident peritumorally. Thus, the integrated model accurately identifies occult invasive risk, addressing a critical limitation inherent in single-region predictive strategies.

MRI provides high soft-tissue resolution and allows comprehensive visualization of organ structures. Dynamic contrast-enhanced MRI (DCE-MRI) effectively highlights lesion enhancement and vascular characteristics, yielding superior lesion conspicuity, which is critical for achieving accurate diagnoses [37]. Given the limited information obtained from single MRI sequences, this study implemented multi-sequence feature fusion. The integration of multi-sequence and multi-region-of-interest (ROI) data synergistically enhanced the prediction performance for MVI. With the continuous development of AI, DL enables automatic extraction of complementary features and facilitates cross-scale integration of multi-phase MRI data, leading to extensive applications in medical imaging [25, 38]. Although DL demonstrates strengths in feature extraction, it generally requires large datasets and extended training periods, unlike conventional machine learning methods [39]. Machine learning algorithms demonstrate superior classification efficiency when trained on limited datasets. Among these, logistic regression (LR) offers balanced classification performance, high interpretability, and computational efficiency，thus avoid the risk of overfitting introduced by complex models [39, 40]. Therefore, this study employed a combined strategy of DL-based feature extraction and LR-based modeling. This approach leverages the strengths of DL in feature mining while utilizing logistic regression to overcome limitations regarding data size and computational demands, balancing model performance with practical utility.

Due to tumor heterogeneity, patients with similar histological characteristics frequently exhibit substantial prognostic differences [41], and single predictive models generally lack stability. To address this issue, clinically relevant imaging factors were incorporated to enhance model robustness. We discovered that a few imaging characteristics, such as tumor size, intrahepatic duct dilatation and arterial edge enhancement ratio were strongly connected to MVI status in ICC, which is consistent with earlier research [13, 14, 16, 17]. In this study, tumor size exhibited marginal significance in univariate analysis (p = 0.05), prompting reflection on the limitations of traditional statistical thresholds (p < 0.05). The statement from the American Statistical Association (ASA) states that p-values serve only as a reference indicator for the strength of evidence, not an absolute standard for determining the presence or absence of an association [42]. Overreliance on “p < 0.05” may lead to misjudgments of clinically important associations.From a clinical perspective, tumor size serves as a core indicator reflecting the stage of tumor progression. This has been validated in multiple studies, which provide strong evidence supporting the clinical value of tumor size [13–17]. Even though the statistical association in this study is borderline, its biological significance should not be overlooked.Tang et al. incorporated the first 24-hour urine output (p = 0.059) and sepsis (p = 0.077) into the final predictive model for acute kidney injury in patients with severe burns, based on the clinical significance of these variables [43]. Our findings indicate that the combined model demonstrated optimal predictive performance, excellent calibration, and strong clinical utility. The SHAP method, a highly effective interpretability tool for machine learning, visually illustrates feature contributions, facilitating clinical translation. In this study, SHAP analysis visualized the weights and influences of the four key predictive features. The DL score exerted a notably higher average impact on the model compared to other features, highlighting its central role in driving model decisions. This finding aligns with clinical insights that the composite score is pivotal in disease stratification, providing quantitative evidence to identify critical predictive factors. Case-specific SHAP analyses further illustrated individual feature contributions, computing personalized prediction probabilities through the Shapley score. This detailed breakdown of feature contributions at the individual level validates model traceability and provides clinicians with concrete insights into the multimodal factors driving disease predictions. However, the SHAP analysis in this study primarily addressed feature importance rankings and single-sample prediction contributions, lacking visualization of high-level semantic features (e.g., lesion boundaries, textures) derived from convolutional layers. This limited the full realization of interpretability goals. Therefore, heatmaps were additionally generated to visualize DL feature contributions. The proposed model enhances the accuracy of preoperative MVI assessment in ICC patients, thereby facilitating personalized and targeted clinical decision-making. It not only improves prediction precision and reduces the assessment duration for MVI status but also optimizes healthcare resource allocation. Moreover, analytical results are intuitively presented through visual heatmaps, clearly identifying high-risk areas for MVI. This visualization significantly improves the positive detection rate in biopsy sampling, preventing false-negative results from blind sampling, thus providing reliable evidence for definitive pathological diagnosis and tailored therapeutic strategies.

Nevertheless, this study has several limitations. First, the retrospective design with a limited sample size may introduce selection bias. Specifically, the small number of MVI-positive cases (63 patients, 31.5%) could limit generalizability and statistical power，particularly regarding the stronger association between tumor size and MVI. Second, imaging data were acquired from multiple MRI scanners, potentially introducing systematic biases due to inherent differences in imaging parameters and reconstruction techniques. Third, feature analysis was limited to the 10-mm peritumoral region, without evaluating other peritumoral extents. The predictive value of alternative peritumoral regions for MVI remains unexplored. Additionally, features were extracted and analyzed solely from the maximum cross-sectional plane of the tumor, potentially overlooking critical three-dimensional heterogeneous characteristics. Lastly, this study did not directly compare the proposed method against traditional radiomics or alternative DL architectures, making comprehensive evaluation of the method’s relative strengths and effectiveness challenging. Future studies will address these limitations by: (1) incorporating prospective multicenter, large-sample data for external validation to systematically assess the generalizability and practical value of the model across diverse clinical settings and imaging protocols; and (2) extending analyses to three-dimensional volumetric regions of interest (VOIs) and performing comprehensive comparisons across multiple peritumoral regions and imaging methodologies to optimize and enhance the performance of the MVI predictive model.

Conclusion

In summary, the combined model integrating clinical imaging features with deep learning features derived from multi-sequence MRI within the tumor and at a 10-millimeter depth around the tumor demonstrated favorable results in predicting the MVI status of ICC patients. This approach may aid in guiding relevant clinical decisions and improving patient prognosis.

Electronic supplementary material

Below is the link to the electronic supplementary material.

Abbreviations

MRI

Magnetic resonance imaging

T1WI

T1-weighted image

T2WI

T2-weighted image

DWI

Diffusion-weighted image

CE-T1WI

Contrast-enhanced T1-weighted imaging

Gd-DTPA

Gadopentate dimeglumine

T2W-FS

T2-weighted fat-suppressed

SI

Signal intensity

ROC

The receiver operating characteristic

ROI

Region of interest

AUC

Areas under the curve

AI

Artificial intelligence

DCA

Decision curve analysis

DL

Deep learning

SHAP

Shapley additive explanation

Grad-CAM

Gradient-weighted Class Activation Mapping

ICC

Intrahepatic cholangiocarcinoma

MVI

Microvascular invasion

AFP

Alpha-fetoprotein

CEA

Carcinoembryonic antigen

CA199

Cancer antigen-19-9

LASSO

Least absolute shrinkage and selection operator

MSE

Mean square error

CV

Cross validation

LR

Logistic regression

Author contributions

C.W. and C.W. conceived the present idea. C.W. wrote the manuscript. C.W., Q.W., X.M., X.Q. and C.Y. were responsible for collecting of the data. C.W., C.W., Q.W. processed and analyzed the data for this study. Y.S. revised and confirmed the manu-script. All authors read and approved the final manuscript.

Funding

This study has received funding by Medical Scientific Research Program of Jiangsu Commission of Health (No.M2021014), Wu Jieping Medical Foundation Clinical Research Special Fund Project (320.6750.2023-11-24), Xuhzou Medical Leading Talent Training Project (No.XWRCHT20210030).

Data availability

The data that support the findings of this study are available from the corresponding author upon reasonable request.

Declarations

Ethics approval and consent to participate

This study was conducted in accordance with the Declaration of Helsinki (as revised in 2013). This retrospective study was approved by the ethical committees of both institutions, Xuzhou Central Hospital (XZXY-LK-20241105-0173) and Zhongshan Hospital of Fudan University (B2021-325 R). Given the retrospective nature of this study, the Ethics Committee granted a waiver of informed consent. All methods were carried out in accordance with relevant guidelines and regulations.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.