Habitat-derived radiomics analysis based on dual lesion for the prediction of microvascular invasion in bifocal hepatocellular carcinoma

Abstract

Purpose

This study aims to predict microvascular invasion (MVI) in bifocal hepatocellular carcinoma (bHCC) by analyzing the habitat and radiomic features of the two tumor lesions, and further to establish a prediction model to estimate the prognosis of bHCC patients.

Methods

183 bHCC patients were enrolled, randomly dividing into a training cohort (n = 146) and a test cohort (n = 37) in an 8:2 ratio. Habitat analysis was performed in two tumors as a whole one to construct the habitat model. The radiomic features were further extracted from the habitat-derived subregions. The logistic regression was used to identify the clinicoradiological variables associated with MVI. The Clinical-Radiological-Habitat (CRH) model combined clinicoradiological and habitat features, while the comprehensive model further integrated habitat-derived radiomics for MVI prediction. Diagnostic performance was assessed using receiver operating characteristic curve analysis, and prognostic analysis was performed using the Kaplan-Meier curves.

Results

The dual tumors in bHCC patients were analyzed as a whole tumor and divided into two habitats (habitat 1 and 2). The habitat model demonstrated AUCs of 0.817 and 0.700 for assessing MVI in the training and testing cohorts, respectively. The CRH model outperformed the habitat model with the AUCs of 0.837 and 0.766 in the training and testing cohorts. The comprehensive model combined habitat-derived radiomics features on arterial phase and T2-weighted imaging with CRH model and achieved excellent diagnostic performance with AUCs of 0.907 and 0.884 in the training and test cohorts, respectively. The model’s clinical relevance was further supported by calibration and decision curve analysis. Additionally, the model showed high predictive accuracy for survival outcomes in risk stratification.

Conclusion

The H_{1 − AP+T2WI}+CRH model hold promise as a non-invasive tool for predicting MVI and overall survival in bHCC patients, offering a valuable approach for preoperative risk assessment.

Clinical trial number

Not applicable.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12880-025-02031-2.

Introduction

Hepatocellular carcinoma (HCC) represents a significant global health issue, ranking as the sixth most common cancer worldwide and the fourth leading cause of cancer-related deaths [1, 2]. Approximately 50% to 75% of patients manifested with multiple lesions within the liver, with bifocal HCC (bHCC) patients accounting for over 50% of cases [3–6]. The surgical resection is considered as the radical treatment for bHCC, whereas the high recurrence rate remains a problem resulting in an overall poor prognosis [7, 8].

Microvascular invasion (MVI) is widely recognized as a significant risk factor for early recurrence and can only be diagnosed by pathology [7–10]. Preoperative prediction of MVI offers valuable insights into patient prognosis [10]. Previous studies have explored the value of traditional imaging features such as unsmooth tumor margin and peritumor arterial enhancement [11] as well as advanced technologies such as radiomics [12] in the prediction of MVI, which demonstrated promising results. However, most of these studies focused on patients with solitary and resectable HCC, limiting the applicability to those with bHCC.

Habitat imaging is an emerging technique that utilizes multiple algorithms to segment tumors into subregions to assess the intra-tumor heterogeneity and the tumor microenvironment [13–16]. Unlike conventional radiomics, which typically analyzes the tumor as a single region of interest, habitat imaging characterizes spatial heterogeneity by automatically segmenting tumors into biologically meaningful subregions, such as invasive margins, hypercellular zones, and necrotic regions [14–16]. These subregions represent distinct aspects of the tumor microenvironment, and subregion-specific feature extraction may improve predictive performance by minimizing the confounding effects of feature averaging across heterogeneous areas [15]. Previous studies have showed the promise of habitat imaging in the prediction of MVI and recurrence-free survival (RFS) in solitary HCC patients, indicating the relationship between intertumoral heterogeneity with MVI and prognosis [17–19]. Compared with solitary HCC, bHCC patients manifested with greater tumor heterogeneity and few studies have explored the association between tumor heterogeneity based on habitat-derived radiomics and MVI status in bHCC [20, 21].

Therefore, this study used habitat imaging and habitat-derived radiomics to characterize the tumor heterogeneity of dual lesions in bifocal HCC patients and to explore its relationship with MVI status. Furthermore, this study developed and validated the model combining habitat-derived radiomics with clinical and conventional radiological variables to predict MVI status of bHCC and prognosis of patients.

Materials and methods

The technical flowchart of this research is depicted in Fig. 1. This retrospective study (B2021-682R) was authorized by Zhongshan hospital’s institutional review committee, following ethical principles outlined in the Declaration of Helsinki. Due to its retrospective nature, written informed consent was not required. We ensured strict confidentiality and protection of patient data.

Fig. 1.

Technical flowchart of this research

Study patients

We retrospectively identified all patients diagnosed with bHCC in MRI using the image archiving and communication system of our academic center from January 2015 to January 2019.

The inclusion criteria for this study were patients who had undergone R0 resection surgery and whose preoperative MRI did not indicate any presence of macrovascular invasion. The total number of patients meeting these criteria was 569. The exclusion criteria were as follows: (a) failure of MRI examination due to patients’ claustrophobia or previous surgical implantation of metal materials (n = 52); (b) either of the two tumors is smaller than 1 cm (n = 61); (c) previous history of treatments before MRI including (hepatectomy, liver transplantation, chemotherapy, radiotherapy, trans-arterial chemoembolization, radiofrequency ablation and systemic therapy) (n = 143); (d) presence of extrahepatic metastasis (n = 15); (e) either of the two lesions pathologically diagnosed as non-HCC after surgery (n = 61); (f) insufficient follow-up data, resulting from participants either refusing to engage in data collection or becoming lost to follow-up due to relocation or supplying incorrect contact information (n = 18); (g) lack of histopathology reports regarding MVI (n = 17); (h) the image quality is compromised by artifacts caused by respiratory motion, as observed by radiologist 1 with three years of experience in liver imaging (n = 8); (i) a significant time lapse (more than 4 weeks) between the MRI scan and the surgical removal of tissue (n = 11).

Ultimately, our study included 183 patients who were randomly assigned to a training set and a validation set in an 8:2 ratio (Figure S1).

Clinicopathological and laboratory data evaluation

All variables were accessed via the electronic medical record system at our institution. The clinical information comprised data on the individual’s gender, age, conditions of HBV or HCV infection and liver cirrhosis. Pathological investigations of the resected specimens were conducted by two qualified pathologists. MVI involves the presence of small blood clots in vessels lined with endothelial cells, including arteries, the hepatic vein, the portal vein, and lymphatic vessels [22]. Specimens for curative hepatectomy were collected along the tumor boundary with neighboring liver tissues, maintaining a 1:1 ratio. Samples were taken at positions corresponding to the 12, 3, 6, and 9 o’clock locations [23]. Pathological MVI was assessed for 2 HCC lesions, and positive features of any lesion were considered positive MVI for the patient. Preoperative laboratory data comprised serum alpha-fetoprotein (AFP) levels, etiology of liver disease, larger tumor diameter (LTD), HBV-DNA load, total tumor diameter (TTD), the ratio of the larger to the smaller tumor diameter (RLSD), total bilirubin (TBIL), direct bilirubin (DBIL), total protein (TP), albumin (ALB), alanine aminotransferase (ALT), aspartate aminotransaminase (AST), alkaline phosphatase (AKP), γ-glutamyl transpeptidase (GGT), total bile acid (TBA), platelet count(PLT), prothrombin time (PT).

Follow-up surveillance

The clinical data for follow-up were obtained through comprehensive review of medical records or via telephone interviews. Specifically, the follow-up was conducted every 3–6 months for the first 2 years after treatment and every 6–12 months thereafter. The primary imaging modalities for monitoring included ultrasound, contrast-enhanced CT, and MRI. In cases of indeterminate findings or suspected metastatic disease, PET/CT was utilized as an adjunctive tool. RFS was defined as the interval from surgery to the first recurrence, metastasis, or last follow-up. Overall survival (OS) was defined as the duration between the surgery and the occurrence of death, the date of the last follow-up, and the study end date on December 31, 2021.

MRI protocol

All patients underwent imaging examination using a 1.5-T MR scanner (MAGNETOM Aera, Siemens Healthcare). Routine liver MR imaging included T1-weighted in-phase and out-of-phase sequences, transverse T2-weighted fast spinecho sequence (T2WI-FS), and diffusion-weighted imaging (DWI) with b values of 800 s/mm². A T1-weighted fat-suppressed sequence was used for dynamic imaging. The intravenous administration of Gadobutrol (Gadavist; Bayer HealthCare) was performed at a rate of 2 mL/s, resulting in a total dose of 0.1 mmol/kg. Upon the arrival of the contrast agent in the ascending aorta, the arterial phase acquisition was automatically initiated. This was followed by the portal venous phase, which occurred between 70 and 90 s, and the delayed phase, which took place between 160 and 180 s. For detailed parameters, refer to Table S1.

MRI data analysis

Two radiologists (with 9, and 20 years of abdominal imaging experience, respectively) evaluated the following qualitative radiological features independently, without access to clinical or pathological data: (1) satellite nodule; (2) hemorrhage in mass; (3) fat in mass; (4) arterial rim enhancement; (5) radiologic capsule; (6) arterial peritumoral enhancement; (7) mosaic architecture; (8) nodule-in-nodule architecture; (9) non-smooth tumor margin; (10) atypical enhancement pattern. Discrepancies among the readers were resolved through consensus discussion following individual image interpretation. The interobserver agreements of MR features are listed in Supplementary Table S2.

The radiologic characteristics of two tumor lesions are assessed, and a positive feature in any lesion is considered a positive radiologic feature for the patient. Each radiologic feature’s definition is given in the Supplementary Material 1.

Tumor segmentation and imaging preprocessing

The sum of the two lesions, which is treated as whole one, was used to analyze the patient’s tumor characteristics. Therefore, this study delineated volumetric region of interest (VOI) of the two lesions in each patient separately. Tumor segmentation was performed using ITK-SNAP software (version 4.0; www.itk-snap.org) by a radiologist 1 (3 years of abdominal imaging analysis experience), and the results were confirmed by a senior radiologist 2 (20 years of abdominal imaging analysis experience). VOIs were manually delineated on six sequences: pre-T1WI, arterial phase (AP), portal venous phase (PVP), delayed phase (DP), T2WI-FS, and diffusion-weighted DWI with b value of 800 s/mm. Additionally, the radiologist 1 reassessed MR images of 30 randomly selected lesions after one month to evaluate intra-observer reproducibility. In order to evaluate inter-observer repeatability, these 30 MR images were also independently re-segmented by a different radiologist 3 (9 years of experience in abdominal imaging analysis).

Image preprocessing consisted of the following steps: (1) N4 Bias Field Correction [the Python-based package “Advanced Normalization Tools”] was used in all MR images to eliminate intensity inhomogeneity correction; (2) In order to achieve the tumor habitat analysis based on multi-sequence information, we used the pre-T1WI sequence as the fixed image and the other sequences as the floating image, and mapped the moving image to the fixed image through the Symmetric normalization [the Python-based package “antspyx”], for the purpose of spatial location consistency between the two sequences; (3) Adaptive normalizer is used to remove voxels with MR image intensity values higher than 99% and lower than 1% in every patient images; (4) Max-min Normalization is used to unify all image intensity values to 0–1.

Habitat quantification and radiomics feature extraction

In this study, the two lesions of bHCC were analyzed collectively as a whole tumor. Using habitat analysis, the tumors were segmented, and radiomic features were subsequently extracted based on habitat subregions and whole tumor region. A K-means unsupervised clustering algorithm based on multiparametric MRI (DWI, T2WI and pre-T1WI) was used to determine the tumor habitats. Utilizing Distortions, Davies-Bouldin Index, and Calinski-Harabasz Index [24–26], we determined 2 subregions to be the optimal number of habitats (Figure S2; Table S3). After completing voxel clustering procedures, voxels within the same cluster were assigned identical colors, producing a clustering map. This map facilitates visualization of spatial clustering patterns and serves as an imaging biomarker for assessing spatial distribution of tumor heterogeneity (Fig. 2).

Fig. 2.

(a–b) Representative habitat images of MVI-positive patient and (c–f) MVI-negative example from the same patient. The blue and green areas represent tumor habitat 1 and 2

First, the habitat segmentations of all patients were quantified with the following formula:

The habitat model was constructed by integrating habitat features (the volume of Habitat 1, Habitat 2, volume share and the total tumor). Following habitat segmentations analysis, radiomics features were extracted from subregions 1 and 2, as well as the entire tumor, across six sequences using uAI Portal (Version: 20230715). The sum of the radiomics characteristics in two lesions represented the patient’s characteristics with the following formula:

After extracting all radiomics features from habitat subregions and the whole tumor region, unstable features with ICC values below 0.75 were excluded [27]. A total of 2250 features were respectively obtained for two habitats and the overall region in each sequence based on the Pyradiomics package with following settings: binWidth = 25, interpolator = BSpline, Sigma = [0.5,1.0,1.5,2.0], resampled Pixel Spacing = [1, 3], and normalizeScale = 1. Details of all radiomics features are provided in Table S4, with parameter justifications and supporting references available in Supplementary Material 2. The multiple-sequence radiomics models were fused the features in the single-sequence radiomics models with an AUC greater than 0.7 [28]. Secondly, radiomics features from multiple sequences were then selected using the Max-Relevance and Min-Redundancy (mRMR) algorithm (method = mutual information and feature number = 20) and the Least Absolute Shrinkage and Selection Operator (LASSO) algorithm (alpha = 1, tolerance = 0.0001 and max iteration = 1000). Finally, logistic regression (LR) and random forest (RF) algorithms were respectively employed to create single-sequence radiomics models in whole tumors and habitat subregions incorporating these features.

Statistical analysis

The Student’s t test was employed for variables that followed a normal distribution, whereas the Mann-Whitney U test was applied for continuous variables that did not follow a normal distribution. The chi-square test was utilized to see if there were statistically significant disparities across qualitative variables. The study employed binary logistic regression analysis to investigate the potential risk factors for MVI with penalty = L2, tolerance = 0.0001, C = 1, class weight = balanced, solver = lbfgs, and max iteration = 100. Clinical and conventional radiologic models were developed utilizing selected risk factors from each domain through logistic regression analysis. Habitat model was combined the habitat features. Furthermore, a comprehensive model was developed incorporating habitat-derived radiomics features, habitat features and clinicoradiological risk factors. The models’ diagnostic ability was assessed using AUC and calibration curves. Decision curve analysis was utilized to compare the net benefit derived from the three different models. The DeLong test was used to compare the predictive performance across models, while both the net reclassification improvement (NRI) and the integrated discrimination improvement (IDI) were computed to evaluate the net gain in predictive accuracy. Statistical analysis was performed with R software (version 4.1.3; R Foundation for Statistical Computing, Vienna, Austria; https://www.R-project.org). All statistical tests were two-sided, and p value lower than 0.05 were deemed statistically significant.

Results

Characteristics of the participants

A total of 183 patients with bifocal HCC who underwent preoperative multiparametric MRI were enrolled and were randomly divided into training set and validation set in a ratio of 8:2. Among the included bHCC patients, 85 (58.22%) and 23 (62.22%) in the training and validation sets were diagnosed with MVI, respectively. The training and test sets had similar characteristics, including age, gender, AFP, etiology of liver disease, LTD, TTD, RLSD, HBV-DNA load, TBIL, DBIL, TP, ALB, ALP, AST, AKP, GGT, TBA, PLT, PT. The detailed patient characteristics are presented in Table 1.

Table 1.

Baseline clinical features of bHCC patients in training and testing cohort

Characteristics	Training cohort			Testing cohort			p value (Train vs Test)
	MVI-(n = 61)	MVI+(n = 85)	p value	MVI-(n = 14)	MVI+(n = 23)	p value
Clinical features
Age (years) *	60[54 ~ 64]	60[52 ~ 63]	0.458	54[50 ~ 58.75]	62[49 ~ 65.5]	0.308	0.395
Gender			0.484			0.544	0.931
Male	56 (91.8)	75 (88.2)		12 (85.7)	22 (95.7)
AFP (ng/mL)			0.100			0.071	0.844
20–400	17 (27.9)	30 (35.3)		4 (28.6)	8 (34.8)
≥400	14 (23.0)	28 (32.9)		1 (7.1)	8 (34.8)
Etiology of liver disease			0.787			0.454	0.518
History of HBV	38 (62.3)	58 (68.2)		9 (64.3)	18 (78.3)
History of HCV	0 (0.0)	2 (2.4)		0 (0.0)	0 (0.0)
Liver Cirrhosis			0.534			0.733	0.183
Positive	32 (52.5)	49 (57.6)		7 (50.0)	9 (39.1)
LTD (cm) *	3.5[2.6 ~ 5]	4[3 ~ 6.5]	0.002	3.4[2.325 ~ 4.5]	5.2[3.75 ~ 7]	0.010	0.219
TTD (cm) *	5.4[4.1 ~ 7]	6.6[5 ~ 10.2]	0.001	4.7[3.325 ~ 7.025]	7.7[5.4 ~ 10.3]	0.004	0.617
RLSD (cm) *	1.846[1.45 ~ 2.89]	2.03[1.39 ~ 3.12]	0.357	2.76[2.025 ~ 3.6]	2[1.5 ~ 3.9]	0.246	0.068
HBV-DNA load (IU/ml)			0.613			0.710	0.711
>104	17 (27.9)	27 (31.8)		3 (21.4)	7 (30.4)
TBIL (µmol/L)			0.799			0.625	0.968
>20.4	5 (8.2)	8 (9.4)		2 (14.3)	2 (8.7)
DBIL (µmol/L)			0.414			0.544	0.149
>6.8	9 (14.8)	17 (20.0)		2 (14.3)	1 (4.3)
TP (g/L)			0.041			1.000	0.281
>65	28 (45.9)	25 (29.4)		6 (42.9)	11 (47.8)
ALB (g/L)			0.014			0.268	0.835
>35	26 (42.6)	20 (23.5)		6 (42.9)	5 (21.7)
ALT (U/L)			0.172			0.390	0.691
>50	10 (16.4)	22 (25.9)		4 (28.6)	3 (13.0)
AST (U/L)			0.123			0.217	0.258
>40	13 (21.3)	28 (32.9)		1 (7.1)	6 (26.1)
AKP (U/L)			0.685			0.275	0.931
>125	7 (11.5)	8 (9.4)		0 (0.0)	3 (13.0)
GGT (U/L)			0.910			0.314	0.239
>60	31 (50.8)	44 (51.8)		4 (28.6)	11 (47.8)
TBA (µmol/L)			0.493			1.000	0.555
>10	21 (34.4)	34 (40.0)		5 (35.7)	7 (30.4)
PLT (*109/L)			0.020			0.039	0.857
>100	27 (44.3)	22 (25.9)		8 (57.1)	5 (21.7)
PT (s)			0.606			0.275	0.956
>13.0	5 (8.2)	4 (4.7)		0 (0.0)	3 (13.0)
Imaging features
Satellite Nodule			0.006			0.376	0.390
Positive	1 (1.6)	13 (15.3)		1 (7.1)	5 (21.7)
Hemorrhage in Mass			0.055			0.191	0.075
Positive	12 (19.7)	29 (34.1)		4 (28.6)	12 (52.2)
Fat in Mass			0.216			0.687	0.292
Positive	20 (32.8)	20 (23.5)		2 (14.3)	5 (21.7)
Arterial Rim Enhancement			0.075			1.000	0.643
Positive	1 (1.6)	9 (10.6)		1 (7.1)	3 (13.0)
Radiological Capsule			0.017			0.017	0.942
Positive	38 (62.3)	36 (42.4)		11 (78.6)	8 (34.8)
Peritumoral Enhancement			0.038			1.000	0.013
Positive	17 (27.9)	38 (44.7)		2 (14.3)	4 (17.4)
Mosaic Architecture			0.006			0.733	0.662
Positive	24 (39.3)	53 (62.4)		7 (50.0)	14 (60.9)
Nodule in Nodule Architecture			0.995			0.378	1.000
Positive	2 (3.3)	4 (4.7)		1 (7.1)	0 (0.0)
Non-Smooth Tumor Margin			0.018			0.434	0.706
Positive	38 (62.3)	68 (80.0)		12 (85.7)	16 (69.6)
Atypical Enhancement Pattern			0.450			0.445	0.148
Positive	9 (14.8)	9 (10.6)		4 (28.6)	4 (17.4)

Data are presented as the number of patients with the percentage in parenthesis, unless otherwise specified

*Data are medians with interquartile ranges in parentheses

Abbreviations: bHCC, bifocal hepatocellular carcinoma; AFP, alpha-fetoprotein; HBV, hepatitis B virus; HCV, hepatitis C virus; TTD, total tumor diameter, RLSD, the ratio of the larger to the smaller tumor diameter; TBIL, total bilirubin; DBIL, direct bilirubin; TP, total protein; ALB, albumin; ALT, alanine aminotransferase; AST, aspartate aminotransaminase; AKP, alkaline phosphatase; GGT, γ-glutamyl transpeptidase; TBA, total bile acid; PLT, platelet count; PT, prothrombin time; MVI, microscopic vascular invasion

In the clinical and radiological features, AFP > 20 ng/mL (p = 0.020, OR = 1.665, 95% CI: 1.089–2.589) and the presence of satellite nodule (p = 0.038, OR = 5.495, 95% CI: 1.325–38.553) were identified as independent risk factors for MVI (Table S5). The clinical model and radiologic model both showed unsatisfactory predictive efficiency (Table 2).

Table 2.

The performance of fusion model using Random Forest

Model	Training cohort						Testing cohort
	AUC (95% CI)	Accuracy	Sensitivity	Specificity	Precision	F1-Score	AUC (95% CI)	Accuracy	Sensitivity	Specificity	Precision	F1-Score
Clinical	0.592(0.504–0.681)	0.603	0.682	0.492	0.652	0.667	0.707(0.549–0.865)	0.676	0.696	0.643	0.762	0.728
Radiologic	0.568(0.527–0.61)	0.500	0.153	0.984	0.929	0.263	0.573(0.462–0.684)	0.486	0.217	0.929	0.833	0.344
Habitat	0.817(0.748–0.886)	0.747	0.741	0.754	0.808	0.773	0.700(0.523–0.878)	0.703	0.783	0.571	0.750	0.766
CRH	0.837(0.774–0.899)	0.767	0.812	0.705	0.793	0.802	0.766(0.61–0.921)	0.757	0.783	0.714	0.818	0.800
AP + T2WI	0.845(0.782–0.908)	0.747	0.671	0.852	0.864	0.755	0.798(0.646–0.950)	0.784	0.870	0.643	0.800	0.834
H1 − AP+T2WI	0.817(0.748–0.885)	0.747	0.659	0.869	0.875	0.752	0.834(0.684–0.984)	0.838	0.913	0.714	0.840	0.875
AP + VP + T2WI	0.827(0.762–0.891)	0.699	0.494	0.984	0.977	0.656	0.812(0.675–0.949)	0.730	0.739	0.714	0.810	0.773
H1 − AP+VP+T2WI	0.823(0.757–0.889)	0.747	0.765	0.721	0.793	0.779	0.736(0.528–0.944)	0.811	0.870	0.714	0.833	0.851
H1 − AP+DP+T2WI	0.841(0.778–0.903)	0.747	0.718	0.787	0.824	0.767	0.814(0.663–0.964)	0.811	1.000	0.500	0.767	0.868
H1 − AP+DP+VP+T2WI	0.819(0.752–0.887)	0.726	0.600	0.902	0.895	0.718	0.734(0.543–0.926)	0.757	0.783	0.714	0.818	0.800
H2 − AP+VP	0.858(0.798–0.918)	0.774	0.694	0.885	0.894	0.781	0.790(0.644–0.937)	0.730	0.609	0.929	0.933	0.737
AP + T2WI + CRH	0.857(0.797–0.918)	0.788	0.894	0.639	0.776	0.831	0.839(0.709–0.968)	0.784	0.696	0.929	0.941	0.800
AP + VP + T2WI + CRH	0.849(0.788–0.911)	0.781	0.847	0.689	0.791	0.818	0.811(0.667–0.954)	0.784	0.739	0.857	0.895	0.810
H1 − AP+T2WI + CRH	0.907(0.858–0.956)	0.884	0.871	0.902	0.925	0.897	0.884(0.773–0.994)	0.838	0.913	0.714	0.840	0.875
H1 − AP+DP+T2WI+CRH	0.852(0.791–0.913)	0.788	0.918	0.607	0.765	0.835	0.851(0.728–0.974)	0.757	0.652	0.929	0.938	0.769
H2 − AP+VP +CRH	0.850(0.789–0.912)	0.774	0.788	0.754	0.817	0.802	0.840(0.711–0.969)	0.784	0.739	0.857	0.895	0.810

Bold values are models with best predictive performance

Abbreviations: CRH, Clinical-Radiologic-Habitat model; AP + T2WI, the radiomics model of total tumor volume in AP and T2WI; H_{1 − AP+T2WI}, the radiomics model of habitat 1 subregion in AP and T2WI; H_2 − AP, the radiomics model of habitat 2 subregion in AP; and so on; AP + T2WI + CRH, the fusion model include the radiomics model of total tumor volume in AP and T2WI and clinical, radiologic, habitat features; H_{1 − AP+T2WI} + CRH, the fusion model include the radiomics model of habitat 1 subregion in AP and T2WI and clinical, radiologic, habitat features

Habitat mapping

The two tumors in bHCC patients were analyzed as a whole tumor and divided into two habitats (habitat 1 and 2). Based on the distinct feature distribution patterns shown in Figure S2 and the intensity distributions of the two tumor subregions across the three imaging sequences summarized in Table S6, we interpret the high signal intensity of Habitat 1 on DWI (b = 800 s/mm²) and T2WI as indicative of elevated cellular density accompanied by increased water content, characteristics typically associated with aggressive, hypercellular tumor regions exhibiting interstitial edema and invasive potential. In contrast, Habitat 2, which demonstrates lower signal intensity on both sequences, likely corresponds to areas of diminished cellularity and moderate water content, potentially reflecting fibrotic stroma, partial necrosis, or other acellular components. Habitat 1 volume is significantly larger in the MVI + group compared to the MVI- group (p = 0.0052), indicating a statistically significant difference between the two groups (Figure S3). Example of representative habitat maps of bHCC with MVI + status was shown in Fig. 2, indicating a higher proportion of volume in habitat 1 correlated more strongly with MVI + status. Based on habitat segmentation result for all patients, the habitat model incorporated five features, including the volume of habitat 1, 2, volume share and total volume. The habitat model demonstrated AUCs of 0.817 (95% CI: 0.748–0.886) and 0.700 (95% CI: 0.523–0.878) for assessing MVI in the training and testing cohorts, respectively. Subsequently, AFP and satellite nodule in imaging, combined with habitat features, were used to constructed the Clinical-Radiological-Habitat (CRH) model. The AUCs for MVI prediction using CRH model were 0.837 (95% CI: 0.774–0.899) for the training cohort and 0.766 (95% CI: 0.61–0.921) for the testing cohort.

The effectiveness of habitat-derived radiomics analysis in single sequences

After extracting all radiomics features from habitat subregions and the whole tumor region, unstable features with ICC values below 0.75 were excluded [27]. Subsequently, features showing the strongest correlation with MVI and the weakest correlation between groups were selected using the mRMR method in each sequence. The final predictive radiomics characteristics for MVI were determined through LASSO algorithm. Through these feature selection process, the number of features associated with MVI for the tumor region and habitat subregions were detailed in Table S7. Single-sequence radiomics models based on habitat subregions and the total tumor volume were then constructed.

The AUCs, accuracy, sensitivity, and specificity of tumor radiomics features and habitat-derived radiomics features in single sequences were presented in Table S8, S9. This study compared efficacy of the single models using the LR and the RF analysis, concluding that RF analysis proved to be more effective than LR analysis. Due to its superior performance, the RF classifier was subsequently chosen for further analysis. Habitat-derived radiomics models outperformed total tumor radiomics models in the AP, DP, and VP sequences using RF (AUC: AP model/ H_1 − AP model/ H_2 − AP model in training cohort = 0.768/0.798/0.818; DP model/ H_1 − DP model in testing cohort = 0.652/0.724; VP model/ H_2 − VP model in testing cohort = 0.710/0.728). Among the habitat-derived radiomics models, the H_2 − AP model demonstrated the best performance, achieving an AUC of 0.818 (95% CI: 0.75–0.886) in the training cohort and an AUC of 0.720 (95% CI: 0.543–0.898) in the testing cohort when using the RF classifier to predict MVI. Among the total tumor radiomics models, the T2WI model showed the best predictive performance using the RF (AUC: 0.82, 95% CI: 0.754–0.887 in the training cohort; AUC: 0.717, 95% CI: 0.53–0.904).

Predictive models development and performance evaluation of fusion models

Firstly, the multiple-sequence radiomics features were fused respectively in the whole tumor and habitat subregions, retaining features with an AUC greater than 0.7 in the single-sequence radiomics models [28]. Secondly, radiomics features from multiple sequences were then selected using the LASSO algorithm. Lastly, RF algorithm was applied to create multi-sequence models. The multi-sequence radiomics models in whole tumor showed superior predictive efficacy compared to the singe-sequence models (AUC: AP + T2WI model/AP model/ T2WI model = 0.845/0.768/0.82; Table 2; Fig. 3). Similarly, the multi-sequence radiomics models in habitat subregions showed better predictive efficacy compared to the singe-sequence models (AUC: H_{1 − AP+DP+T2WI} model/H_{1 − AP} model/H_{1 − DP} model/ H_{1 − T2WI} model = 0.841/0.798/0.797/0.786; Table 2; Fig. 3). Habitat 1- derived radiomics models in AP and T2WI sequences combined with CRH model had best predictive performance (AUC: H_{1 − AP+T2WI}+CRH model/ AP + T2WI + CRH model/ H_{1 − AP+T2WI} model = 0.907/ 0.857/ 0.817).

Fig. 3.

AUCs of habitat-derived radiomics models for assessing MVI in single sequences (a: training cohort, b: testing cohort) and multiple sequences (c: training cohort, d: testing cohort). The radar chart of fusion models for MVI prediction (e: in training cohort, f: testing cohort)

The calibration curve revealed the prediction results of each fusion model closely aligned with the actual outcomes, as the curves are positioned around the main diagonal (Fig. 4a-b). In the test set, the p-values of the Hosmer–Lemeshow test for the AP + T2WI + CRH, AP + VP + T2WI + CRH, H_2 -AP+VP+CRH, H_{1 -AP+DP+T2WI} + CRH, and H_{1 − AP+T2WI}+CRH models were 0.025, 0.020, 0.357, 0.067, and 0.215, respectively, with corresponding Brier scores of 0.026, 0.012, 0.012, 0.036, and 0.002. The DCA curve illustrated that the H_{1 − AP+T2WI}+CRH model, with classification thresholds of 0.4 to 0.65, provided greater clinical benefits compared to other models (Fig. 4c-d). Additionally, the confusion matrix of the H_{1 − AP+T2WI} + CRH models was shown in Fig. 5, which also illustrated the satisfactory effect of the predictive model.

Fig. 4.

Evaluation and verification of the fusion models. (a) Calibration curves of the fusion models in the training cohort and (b) in the training cohort. (c) Decision curves of the fusion models in the training cohort and (d) in the training cohort. The grey line and horizontal black line represent the assumption that all and no patients have MVI, respectively

Fig. 5.

(a) Confusion matrix of the H_{1 − AP+T2WI} + CRH model in training cohort and (b) in test cohort

The predictive performance of the CRH model showed a dramatic increase compared to that of Clinical model (NRI > 0, p = 0.001), and Radiological model (NRI > 0, p < 0.001). Similarly, significant improvements were observed between the H_{1 − AP+T2WI} + CRH model and the H_{1 − AP+T2WI} model (IDI > 0, p < 0.001), as well as between the H_{1 − AP+T2WI} + CRH model and the CHR model (IDI > 0, p < 0.001) (Table 3). The DeLong test confirmed that both the H_{1 − AP+T2WI} + CRH and CRH models outperformed other models (p -training cohort < 0.001) (Table S10).

Table 3.

(a) Net reclassification indexes of diverse combination models, (b) Integrated discrimination improvement test of diverse combination models

(a)
Compared models	Basic model	NRI | p value
		Training cohort	Testing cohort
H1 − AP+T2WI +CRH	H1 − AP+T2WI	0.310 | <0.001	0.146 | 0.247
H1 − AP+T2WI +CRH	CRH	0.537 | <0.001	0.348 | 0.059
CRH	Clinical	0.343 | 0.001	0.158 | 0.405
CRH	Radiologic	0.380 | <0.001	0.351 | 0.020
CRH	Habitat	0.138 | 0.042	0.230 | 0.156
H1 − AP+T2WI	H1 − AP	0.101 | 0.174	0.242 | 0.088
H1 − AP+T2WI	H1 − T2WI	0.210 | 0.005	0.230 | 0.156

(b)
Compared models	Basic model	IDI | p value
		Training cohort	Testing cohort
H1 − AP+T2WI +CRH	H1 − AP+T2WI	0.155 | <0.001	0.041| 0.296
H1 − AP+T2WI +CRH	CRH	0.279 | <0.001	0.250 | 0.003
CRH	Clinical	0.104 | <0.001	0.022 | 0.611
CRH	Radiologic	0.080 | 0.001	0.023 | 0.647
CRH	Habitat	0.038| 0.001	0.009| 0.684
H1 − AP+T2WI	H1 − AP	0.029 | 0.284	0.089 | 0.160
H1 − AP+T2WI	H1 − T2WI	0.018 | 0.519	0.112 | 0.050

Abbreviations: CRH, Clinical-Radiologic-Habitat; AP + T2WI + CRH, the fusion model include the radiomics model of total tumor volume in AP and T2WI and clinical, radiologic, habitat features; H_{1 − AP+T2WI} + CRH, the fusion model include the radiomics model of habitat 1 subregion in AP and T2WI and clinical, radiologic, habitat features

Clinical outcome evaluation

Until December 2021, 183 patients had completed follow-up, with the median RFS of 29 months (1–82 months) and median OS of 48 months (7–82 months). The 1-, 3-, and 5-year RFS rates were 65.89%, 37.45%, and 31.36%, respectively. The 1-, 3-, and 5-year OS rates were 95.08%, 81.42%, and 73.77%, respectively.

There was a significant difference in RFS between MVI-positive and MVI-negative patients (median RFS, 14 vs. 42 months; p < 0.0001). Similarly, based on our optimal model, patients with model predicted MVI had significantly shorter RFS (14 months vs. 40 months, p < 0.0001) than those without (Figure S4).

In pathology, for MVI-positive bHCC compared to MVI-negative bHCC, the median OS was 43 months versus 54 months (p < 0.001, shown in Fig. 6a). According to the H_{1 − AP+T2WI}+CRH model, patients were stratified into MVI high (cutoff value > 0.625) and low-risk groups. Similarly, for MVI-pre-positive HCC compared to MVI-pre-negative HCC, the median OS was 45 months versus 53 months (p < 0.001, shown in Fig. 6b). Predictive-MVI and Nodule in Nodule Architecture in imaging were independent risk variables for the outcome, according to the multivariate Cox regression results (Table 4), with HRs of 5.646 (95% CI: 2.1-15.18) and 4.476 (95% CI: 1.277–15.69), respectively. Figure 6c displayed the OS nomogram based predictive-MVI subgroup for identifying the overall survival risk, and the high-risk group exhibited a lower OS compared to the low-risk group (Fig. 6d).

Fig. 6.

(a) Kaplan-Meier curves of overall survival in histological MVI subgroup and (b) in predictive MVI subgroup. (c) Nomogram based predictive-MVI subgroup for identifying the overall survival risk and (d) Kaplan-Meier curves

Table 4.

Variables associated with overall survival according to the Cox proportional hazards model

Characteristics	Univariate analysis		Multivariate analysis (histological MVI subgroup)		Multivariate analysis (predicted-MVI subgroup)
	P value	HR (95% CI)	P value	HR (95% CI)	P value	HR (95% CI)
MVI	< 0.001	6.938(3.605–13.352)	0.001	5.805 (2.168–15.52)
MV-Pre	< 0.001	3.455(2.115–5.644)			0.044	2.217 (1.012–4.856)
Gender	0.163	0.669(0.38–1.177)
AFP	< 0.001	1.688(1.328–2.147)	0.067	1.452 (0.9655–2.181)	0.060	1.469 (0.9862–2.188)
Etiology of liver disease	0.002	2.103(1.325–3.337)	0.921	1.045 (0.4293–2.543)	0.84	1.094 (0.4483–2.673)
Liver Cirrhosis	0.103	1.397(0.935–2.089)
Age	0.550	1.062(0.872–1.294)	0.596	1.008 (0.9795–1.037)	0.565	2.009 (0.858–4.704)
LTD	< 0.001	1.413(1.23–1.624)	0.289	0.804 (0.5863–1.104)	0.338	0.821 (0.6027–1.118)
TTD	< 0.001	1.468(1.251–1.721)	0.327	1.174 (0.8423–1.635)	0.344	1.171 (0.8413–1.629)
RLSD	< 0.001	1.419(1.189–1.693)	0.119	1.310 (0.9317–1.840)	0.187	1.263 (0.8893–1.794)
TBIL	0.822	0.921(0.447–1.897)
DBIL	0.130	0.627(0.343–1.147)
TP	0.109	0.71(0.467–1.08)
ALB	< 0.001	0.356(0.202–0.628)	0.411	0.6140 (0.2295–1.644)	0.393	0.589 (0.231–1.502)
ALT	0.500	0.845(0.518–1.379)
AST	< 0.001	2.175(1.466–3.229)	0.18	1.568 (0.8040–3.059)	0.133	1.644 (0.8596–3.145)
AKP	0.735	1.119(0.582–2.152)			0.888	1.064 (0.4296–2.634)
GGT	0.335	1.211(0.82–1.788)
TBA	0.242	1.267(0.852–1.883)
PLT	0.002	0.449(0.27–0.748)	0.8	1.115 (0.4500–2.760)
PT	0.831	0.914(0.4-2.087)
Satellite Nodule	0.000	2.391(1.511–3.782)	0.215	1.667 (0.7107–3.911)	0.183	1.745 (0.7536–4.036)
Arterial Rim Enhancement	0.005	2.157(1.261–3.687)	0.99	0.994 (0.3581–2.758)	0.821	1.125 (0.4062–3.113)
Radiological Capsule	0.004	0.553(0.368–0.83)	0.776	0.88 (0.3915–1.977)	0.639	0.814 (0.3671–1.804)
Peritumoral Enhancement	0.850	1.04(0.693–1.562)
Mosaic Architecture	0.002	1.925(1.273–2.909)	0.664	0.852 (0.4417–1.643)	0.893	0.952 (0.5112–1.771)
Nodule in Nodule Architecture	0.018	2.402(1.163–4.961)	0.024	4.056 (1.21–13.6)	0.027	3.92 (1.167–13.18)
Non-Smooth Tumor Margin	0.005	2.159(1.263–3.688)	0.24	1.653 (0.6696-4.08)	0.257	1.628 (0.7049–3.755)
Atypical Enhancement Pattern	0.173	0.635(0.33–1.22)

The COX proportional hazards model was established on the basis that the data met the proportional hazards assumption. By Schoenfeld residual method, we found that the p values of HBV-DNA Load, Hemorrhage in Mass, and Fat in Mass in radiologic were lower than 0.05, which did not meet the proportional hazards assumption. Therefore, the above features were removed before univariable and multivariable COX analysis

Discussion

This research developed a novel approach utilizing habitat analysis for noninvasively predicting MVI in patients with bHCC. The dual tumor lesions were analyzed as a whole one and separated into 2 subregions (habitat 1 and 2). The habitat model demonstrated AUCs of 0.817 and 0.700 for assessing MVI in the training and testing cohorts, respectively. Further, the CRH model that combined the clinicoradiological features enhanced the predictive performance with the AUCs of 0.837 and 0.766 in the training and testing cohorts, respectively. Based on the habitat subregions, radiomic features were extracted from two habitats and whole tumor region in single and multiple sequences respectively. The results showed that when integrating habitat 1-derived radiomics in AP and T2WI, the comprehensive model (H_{1 − AP+T2WI}+CRH model) demonstrated the best performance in predicting MVI status in bHCC patients, with AUC values of 0.907 and 0.884 in the training and test cohort.

HCC is characterized by the high degree of heterogeneity, and habitat imaging has emerged as an innovative technique tool to characterize tumor heterogeneity [9, 13, 18, 29]. In the prior study, Zhang et al. [19] demonstrated that habitat analysis of HCC can be used to predict MVI and tumor recurrence, highlighting a strong association between tumor heterogeneity and MVI. However, their study was conducted in patients with solitary HCC. On the patient level, bHCC exhibits greater tumor heterogeneity compared to solitary HCC [31]. Therefore, this study utilized habitat imaging to analyze both lesions as a whole and extracted radiomic features from habitat subregions. By leveraging high-dimensional data, we comprehensively assessed tumor heterogeneity at the microscopic scale and compared the efficy of habitat-derived radiomics with those extracted at the whole-tumor level, demonstrating satisfactory predictive performance. Importantly, we used ICC analysis, LASSO, and mRMR methods during dual-lesion radiomics feature extraction to reduce overfitting and multicollinearity, following radiomics study standards [27].

Except from the habitat-derived radiomics, elevated AFP levels and satellite nodules on imaging were associated with MVI, aligning with prior findings [30–33]. In our study, we comprehensively analyzed the two lesions in bHCC as a unified entity to characterize tumor habitat imaging features, addressing the limitation of previous models that were not applicable to bHCC. The larger Habitat 1 region was strongly associated with MVI positivity, likely reflecting a combination of high cell density (diffusion restriction on DWI with b = 800 s/mm²) and elevated water content (T2WI hyperintensity). These features correspond to cell-rich, edematous tumor regions and may indicate more aggressive biological behavior. Importantly, radiomics features extracted from Habitat 1 outperformed those derived from the entire tumor, suggesting a close association between tumor heterogeneity and MVI. In contrast, Habitat 2, which may represent regions with low tumor cellularity, fibrosis, and partial necrosis, provides more accurate and efficient imaging information by discarding redundant features. Utilizing multiparametric MRI, the habitat framework integrates valuable information, offering insights into the overall characteristics of bifocal tumors. The inclusion of H_{1 − AP} and H_{1− T2WI} features further enhanced model performance, likely due to arterial phase hyperenhancement, a key diagnostic feature according to LI-RADS [34], and the morphological information captured by T2WI sequences.

This study has several limitations. First, although it serves primarily as an exploratory investigation of the relationship between habitat reflecting tumor heterogeneity and MVI in patients with bHCC, including only surgical candidates may introduce selection bias and limit the generalizability of our findings to broader HCC populations, ideally those with advanced-stage disease ineligible for resection. Second, the study was conducted at a single center with a relatively small sample size. Validation in larger, multi-center external cohorts, particularly using temporally or geographically independent datasets, is essential to confirm the robustness and wider applicability of our findings. Third, because this study employed habitat-based radiomics to predict MVI at the patient level rather than evaluating MVI for each individual lesion, a lesion-level comparative analysis of habitat features could provide valuable biological insights. Fourth, spatially matched correlation between preoperative imaging and histopathology was not feasible within the scope of this retrospective design, limiting the definitive biological validation of the identified habitat regions. Prospective studies incorporating systematic radiologic–pathologic correlation are therefore warranted to confirm the histopathologic basis of the proposed imaging-derived habitats.

Conclusion

Our study introduces a novel habitat-derived radiomics approach for the non-invasive, preoperative prediction of MVI and clinical outcomes in patients with bHCC. The findings reveal that the CRH model outperforms the habitat model in predictive accuracy for MVI. Furthermore, the H_{1 − AP+T2WI}+CRH model demonstrates robust performance in both the training and test cohorts, suggesting its potential as a tool for personalized clinical management.

Supplementary Information

Below is the link to the electronic supplementary material.

Abbreviations

bHCC

Bifocal hepatocellular carcinoma

MVI

Microvascular invasion

mRMR

Max-Relevance and Min-Redundancy

LASSO

Least Absolute Shrinkage and Selection Operator

AUC

Area under the curve

ROC

Receiver operating characteristic curve

MRI

Magnetic resonance imaging

HCC

Hepatocellular carcinoma

mHCC

Multifocal hepatocellular carcinoma

ITH

Intra-tumoral heterogeneity

AFP

Alpha-fetoprotein

HBV

Hepatitis B virus

RLSD

Larger to the smaller tumor diameter

Author contributions

All authors are aware of and agree to the submission and that they have all contributed to the work described sufficiently to be named as authors. Xi Jia and Fei Wu were responsible for manuscript writing and statistical analysis. Jing Liu contributed to clinical data collection. Fang Wang, Yuyao Xiao and Dijia Wu helped for image analysis technique. Chun Yang helped for improvement of article structure. Mengsu Zeng designed the study, and revised the manuscript.

Funding

This work was supported by National Natural Science Foundation of China (Grant number [82371923] and [82171897]), Shanghai Municipal Health Commission (Grant number [202240152]), China National Key R&D Program (Grant number [2022YFC2401605]), Scientific Research Development Plan of SHDC and UNITED IMAGING and Science and Technology Commission of Shanghai Municipality (Grant number [23Y11907400]).

Data availability

The datasets used and/or analyzed during the current study are available from the corresponding author on reasonable request.

Declarations

Ethics approval and consent to participate

This retrospective study (B2021-682R) was authorized by Zhongshan hospital’s institutional review committee, following ethical principles outlined in the Declaration of Helsinki. Due to its retrospective nature, written informed consent was not required. We ensured strict confidentiality and protection of patient data.

Consent to participate

Patient informed consent was waived.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.