A habitat radiomics model based on contrast-enhanced MRI for predicting early treatment response to hepatic arterial infusion chemotherapy in patients with unresectable hepatocellular carcinoma

Abstract

Purpose

To develop a contrast-enhanced Magnetic Resonance Imaging (CEMRI)-based habitat radiomics model for predicting early treatment response to hepatic artery infusion chemotherapy with fluorouracil, leucovorin, and oxaliplatin (HAIC-FOLFOX) in patients with unresectable hepatocellular carcinoma (HCC) and elucidate the underlying biological mechanisms.

Methods

Among 120 HCC patients who underwent HAIC treatment, habitat features were extracted by applying clustering algorithms to preoperative CEMRI to delineate intratumoral subregions with distinct enhancement characteristics. Least absolute shrinkage and selection operator (LASSO) and logistic regression were employed for feature selection to construct habitat, conventional radiomics, clinical, and combined models. Internal validation was performed using 1000 bootstrap resamples. In a separate cohort of 107 surgically resected HCC patients, the habitat model was applied for risk stratification, and correlations between habitat features and pathomorphological characteristics, as well as immunohistochemical (IHC) markers, were investigated.

Results

MRI images were categorized into three distinct habitats, and a predictive model was built from their proportional distribution. The habitat radiomics model achieved an area under the curve (AUC) of 0.868 (95% confidence interval (CI): 0.748–0.976), outperforming both conventional radiomics (AUC 0.849, 95% CI: 0.719–0.954) and clinical models (AUC 0.653, 95% CI: 0.497–0.802). The combined clinical-habitat model reached the highest AUC of 0.901 (95% CI: 0.795–0.989, P < 0.05). In the surgical cohort, low-risk habitat patients exhibited increased tumor necrosis/stromal components (elevated IntensityMin) and better differentiation (reduced CurvMean) (P < 0.05). Immunohistochemistry revealed higher microvessel density (CD34) and lower cancer stem cell marker expression (CK19, Glypican-3) in the low-risk group (P < 0.05).

Conclusion

The CEMRI habitat radiomics model accurately predicts early HAIC treatment response, with risk stratification significantly correlated with pathomorphological and molecular characteristics.

Introduction

Hepatocellular carcinoma (HCC) is the dominant pathological type of primary liver cancer, accounting for about 90% of cases [1–3], and ranking as the 6th most common malignancy and the 3rd leading cause of cancer-related death worldwide [4, 5]. Notably, approximately 72% of cases occur in Asia, with China alone accounting for about 50% of the global burden [6, 7]. Most patients are diagnosed at intermediate to advanced stages, missing the window for curative treatment options like surgical resection or liver transplantation [8–12].

For unresectable HCC, hepatic arterial infusion chemotherapy (HAIC) based on the oxaliplatin, leucovorin, and fluorouracil (FOLFOX) regimen has emerged as a pivotal treatment strategy in Asia [13–17]. Recent studies have demonstrated that combination therapies of HAIC with targeted agents and immune checkpoint inhibitors achieve promising objective response rates (ORR) in advanced HCC patients with portal vein tumor thrombus (PVTT) [12, 18–20]. However, due to the high-level heterogeneity of individual response [21, 22], approximately 30%-40% of patients still respond poorly to HAIC, not only failing to derive benefit but potentially experiencing diminished quality of life due to treatment delays and increased risk of complications [18, 19, 23, 24]. More importantly, liver cirrhosis is the most common comorbidity in HCC patients, and hepatic functional reserve critically influences survival, as outcomes depend on both tumor progression and iatrogenic liver impairment [15, 25]. Although FOLFOX-HAIC exhibits a more favorable liver safety profile than transcatheter arterial chemoembolization (TACE), repeated ineffective cycles can still induce chemotherapy-related liver injury. A Phase III trial demonstrated that FOLFOX-HAIC causes grade 3–4 Alanine aminotransferase (ALT) and Aspartate aminotransferase (AST) elevations in 8% and 18% of HCC patients, respectively, and this cumulative chemotoxic damage further compromises hepatic reserve in cirrhotic patients [15]. Such liver dysfunction not only reduces treatment tolerance but also independently worsens survival. Consequently, identifying HAIC non-responders before treatment initiation is clinically urgent to avoid futile therapy and maximally preserve hepatic function. The precise selection of potential HAIC beneficiaries is therefore essential for optimizing clinical decisions and improving the overall prognosis of HCC patients.

HCC is characterized by significant tumor heterogeneity and inherent resistance [26, 27], which poses a challenge to the predictive reliability of existing biomarkers (including serological markers, clinical staging, and conventional radiomic features), which often lack clear biological interpretability [27, 28]. In recent years, habitat radiomics has emerged as a cutting-edge technology for dissecting tumor spatial heterogeneity, demonstrating tremendous potential in advancing precision oncology [29–32]. By employing unsupervised clustering algorithms to deconstruct tumors into physiologically distinct subregions, this technology quantifies their spatial distribution and interactions, thereby enabling refined characterization of tumor heterogeneity [33]. Recently, this approach has achieved breakthrough applications in HCC, including postoperative recurrence risk prediction, microvascular invasion assessment, and pathological grading determination [31, 33, 34]. Our previous study has already demonstrated that Computed Tomography (CT)-based habitat analysis effectively predicts HAIC efficacy in HCC [41]. Notably, Magnetic Resonance Imaging (MRI) offers distinct advantages over CT. Its superior soft tissue resolution enables precise tumor segmentation, while multi-parametric functional-morphological assessment captures biological tumor traits that predict response to HAIC. Moreover, MRI involves no ionizing radiation and is far more sensitive in distinguishing residual viable tumor from treatment necrosis. Nevertheless, whether contrast-enhanced MRI (CEMRI)-based habitat models can effectively predict HAIC therapeutic efficacy remains an unexplored research gap in the field and warrants urgent investigation.

This study aims to develop a CEMRI-based habitat radiomics model for predicting early treatment response to HAIC-FOLFOX regimen in patients with unresectable HCC, and to systematically interpret its potential biological mechanisms of this model through integrated pathomic and immunohistochemistry analyses from a multi-omics perspective.

Materials and methods

Study design and patient cohorts

This retrospective multicenter study was approved by the Ethics Committees of the three participating centers. As illustrated in Fig. 1, this study retrospectively recruited 1276 HCC patients who received HAIC-FOLFOX treatment between January 2022 and December 2024 from three medical centers, the First Affiliated Hospital of Chongqing Medical University (CQH, n = 596), Second Affiliated Hospital of the Army Medical University (XQH, n = 415), and Army Medical University Daping Hospital (DPH, n = 265). Additionally, 493 HCC patients who underwent surgical resection at DPH between January 2023 and December 2024 were also retrospectively screened. All patients were diagnosed with HCC according to the American Association for the Study of Liver Disease (AASLD) guidelines and met the following criteria: (1) Received either HAIC treatment or surgical resection for HCC; (2) Eastern Cooperative Oncology Group (ECOG) performance status of 0–1 and Child-Pugh class A or B; (3) Underwent CEMRI within 2 weeks before treatment; (4) Underwent follow-up contrast-enhanced CT or MRI within 2 months after treatment for response evaluation; (5) No radiotherapy, TACE, or ablation therapy before treatment; (6) Absence of massive ascites; (7) No severe hypersplenism (white blood cell count < 2 × 10⁹/L, platelet count < 50 × 10⁹/L); (8) No severe cardiac, pulmonary, or renal dysfunction; (9) Without other malignant tumors; (10) Follow-up duration exceeding 2 months with complete clinical and imaging data; (11) high image quality. Ultimately, 120 patients from the HAIC-treated cohort and 107 patients from the surgical resection cohort were enrolled for subsequent analysis. Furthermore, Cirrhosis was diagnosed based on clinical, laboratory, and imaging features. The HAIC-treated cohort comprised 98 patients with cirrhosis and 22 without. The pipeline of the study design is shown in Fig. 2.

Fig. 1.

Flow chart of patients included and excluded in this study

Fig. 2.

Technical roadmap of this study. (A) The Hepatic Arterial Infusion Chemotherapy (HAIC) patient cohort and surgical resection patient cohort were collected. (B) Tumor lesions were segmented from Contrast-Enhanced Magnetic Resonance Imaging (CEMRI). (C) Followed by radiomics feature extraction and K-means clustering to determine the optimal cluster number. Features were then screened using intraclass correlation coefficients (ICCs), correlation analysis, and the least absolute shrinkage and selection operator (LASSO). A habitat radiomics model was constructed with logistic regression and evaluated. (D) The underlying biological mechanisms of habitat features were interpreted through pathomics

Follow-up and treatment response evaluation

All patients were followed up until December 31, 2024. Treatment response was assessed according to the modified Response Evaluation Criteria in Solid Tumors (mRECIST) [34, 35] using contrast-enhanced CT or MRI 1–2 months after two cycles of HAIC therapy. Patients were categorized into an objective response (OR: complete response [CR] + partial response [PR]) group and a non-response (nOR: stable disease [SD] + progressive disease [PD]) group. Image evaluations were independently performed by two abdominal radiologists, each with over 5 years of experience. Any discrepancies were resolved through consensus discussion. Consistency between contrast-enhanced CT and contrast-enhanced MRI for response assessment was verified in 18 patients who underwent both examinations within a short interval, with a Cohen’s kappa coefficient of 0.82 (P < 0.001).

MRI image acquisition

CEMRI examinations were performed using 1.5T or 3.0T MR scanners (Siemens Verio). The protocol included coronal T1-weighted imaging (T1WI) and contrast-enhanced T1WI. A bolus of 10 mL gadoxetic acid (Primovist) followed by 20 mL saline was injected via the antecubital vein at a rate of 1–2 mL/s. Arterial phase (AP) scanning was triggered using a bolus-tracking technique, followed by portal venous phase (VP), delayed phase (DP), and hepatobiliary phase (HBP) scans at 35 s, 125 s, and 20 min post-injection, respectively. Key sequence parameters were: TR 15 ms, TE 1.31 ms, slice thickness 4 mm, matrix 256 × 205, NEX = 1.

Image analysis pipeline

The habitat imaging analysis comprised four main steps: image preprocessing, tumor segmentation, habitat encoding, and habitat feature extraction.

Image preprocessing

All MRI images were resampled to an isotropic voxel space of 1 × 1 × 1 mm³ and discretized into 256 Gy levels to minimize the influence of resolution and intensity variations on subsequent analysis. A pretrained nnUNet model (Task003_Liver) was then employed for automatic liver segmentation, and the results were manually verified and corrected by two experienced radiologists. Finally, multi-phase image registration was performed using Elastix software. Using the AP images as the reference, B-spline transformation was applied to non-linearly register the other phase images [36], ensuring spatial consistency across different contrast phases.

Tumor segmentation

On the registered arterial phase T1WI images, two abdominal radiologists independently manually delineated the three-dimensional tumor volume of interest (VOI) using 3D Slicer software (v5.2.2). For cases with multiple nodules, the two largest lesions (both > 2 cm in diameter) were selected as target lesions.

Habitat encoding

Based on the registered multi-phase images, the arterial phase enhancement (AN) map (Fig. 3A) and hepatobiliary phase enhancement (HN) map (Fig. 3B) were calculated, respectively, using the following formulas:

Fig. 3.

Features used to encode habitats and their distribution. (A) Arterial enhancement (AN) parameter map. (B) Hepatobiliary phase enhancement (HN) parameter map. (C) Elbow method plot. (D) Scatter plot of habitat feature distribution. (E) Distribution of AN and HN values across different habitats. (F) Proportion of different habitats. (G) Proportion of different habitats in AN. (H) Proportion of different habitats in HN

Where SI represents signal intensity, and Pre, AP, and HBP denote the pre-contrast, arterial, and hepatobiliary phases, respectively. After extracting the AN and HN values for all voxels within the tumor VOI, unsupervised clustering was performed using the K-means algorithm (Python scikit-learn package). The number of candidate cluster numbers ranged from 2 to 15 and was evaluated using both the elbow method and the silhouette coefficient to determine the optimal number.

Habitat feature extraction

Based on the generated habitat label maps, a Multi-region Spatial Interaction (MSI) matrix was constructed to quantify the spatial relationships between different habitats. This matrix computed the co-occurrence frequency of habitat types for each voxel and its neighboring voxels. A total of 34 features were subsequently extracted from the MSI matrix, comprising 30 first-order statistical features (reflecting the quantity, proportion, and distribution of each habitat) and 4 s-order texture features (characterizing the spatial interaction patterns between habitats) [18]. Furthermore, the quantitative characteristics of the habitat voxels were calculated, including the habitat volume and its proportional composition, which collectively reflect the extent of intratumoral spatial heterogeneity.

The stability of the extracted features was evaluated by performing intra- and interreader reproducibility assessments. For this purpose, two radiologists (L.G.H. and W.H., with 3 and 10 years of abdominal MRI experience, respectively) manually delineated the VOIs of the tumor using 3Dslicer software. Intraclass correlation coefficients (ICCs) were used to assess reproducibility. Habitat features with intra- and inter-reader ICCs > 0.8 were chosen for subsequent analysis. Moreover, features were subjected to t-tests or Mann-Whitney U tests based on treatment response classification, with those achieving a p-value < 0.05 retained. Spearman correlation coefficients (r) were calculated for pairwise combinations of the remaining features. From feature pairs with r > 0.95, the feature demonstrating the most significant p-value was preserved. Finally, the Least Absolute Shrinkage and Selection Operator (LASSO) regression with 10-fold cross-validation was employed for dimensionality reduction of the retained features.

Conventional radiomics feature extraction

Using PyRadiomics (v3.1.0), 107 conventional radiomics features were extracted from the tumor VOI on arterial phase T1WI images. These features primarily included the following four categories: First-order statistics (18 features), describing the distribution characteristics of voxel intensities within the tumor VOI; Shape-based features (14 features), quantifying the three-dimensional geometry of the tumor; Textural features (54 features), comprising Gray Level Co-occurrence Matrix (GLCM, 24 features), Gray Level Dependence Matrix (GLDM, 14 features), and Gray Level Run Length Matrix (GLRLM, 16 features) features; Higher-order features (21 features), including Gray Level Size Zone Matrix (GLSZM, 16 features) and Neighboring Gray Tone Difference Matrix (NGTDM, 5 features) features.

Model construction and evaluation

Screened imaging features were utilized to construct the habitat radiomics model using binary logistic regression. Multivariate logistic regression was employed to identify independent clinical predictors of treatment response. Subsequently, an integrated prediction model was developed by combining the habitat model with clinically relevant variables identified through multivariate analysis.

The discriminatory performance of the models was evaluated using receiver operating characteristic (ROC) curves, calibration plots, and confusion matrices. Model comparisons were performed using DeLong’s test. Internal validation was conducted via bootstrap resampling with 1000 iterations, and estimated the AUC and its 95% confidence interval (CI). Additionally, model calibration analysis was performed. Finally, decision curve analysis (DCA) was employed to evaluate the clinical utility of the models.

Pathomics and immunohistochemical analysis

Hematoxylin and eosin (H&E) and immunohistochemical (IHC) staining were performed on surgical specimens from the 107 patients in the surgical resection cohort. CellProfiler software was used to extract 18 nuclear and cytoplasmic morphological features from whole-slide digital images. IHC expression was quantified using the H-score method: H-score = Σ (percentage of weakly positive cells × 1 + percentage of moderately positive cells × 2 + percentage of strongly positive cells × 3). Assessments were independently performed by two senior pathologists, each with over 3 years of experience.

Statistical analysis

Computational implementation and visualization were conducted using Python 3.7.6 and R 4.2.3. Statistical analyses were performed with IBM SPSS Statistics version 26.0. The Shapiro-Wilk test was used to evaluate whether the continuous data conformed to a normal distribution. Continuous variables with normal distribution were expressed as mean ± standard deviation and compared using the Student’s t-test, while those with non-normal distribution were expressed as median (interquartile range) and compared using the Mann-Whitney U test. Categorical variables were presented as frequency (percentage) and compared using the Pearson’s chi-square test or Fisher’s exact test, as appropriate. For all analyses, statistical significance was set at P < 0.05.

Results

Patient baseline characteristics

As illustrated in Figs. 1 and 1276 HCC patients who received HAIC-FOLFOX treatment were initially screened. After applying strict inclusion and exclusion criteria, 683 were excluded due to prior pretreatment, 95 for concurrent malignancies, 168 for severe cardiac, pulmonary, or renal dysfunction, 89 for poor image quality, and 121 for insufficient follow-up data. Consequently, 120 patients were enrolled in the HAIC cohort. The cohort comprised 85 responders (R) and 35 non-responders (NR) with comparable mean ages. Platelet counts were significantly higher in the response group (P < 0.05). The proportion of patients with alpha-fetoprotein (AFP) > 400 µg/L was significantly higher in responders (P < 0.05). No other significant differences in clinical characteristics were observed (P > 0.05, Table 1). In addition, to further explore the potential influencing factors of treatment response in HCC patients receiving HAIC-FOLFOX, we employed multivariate logistic regression analysis. Platelets (Odds Ratio (OR) = 1.01, P = 0.038) and AFP > 400 µg/L (OR = 2.63, P = 0.022) were still found to be independent predictors (Table 2).

Table 1.

Baseline characteristics of HAIC cohort patients

Variables	Total (n = 120)	NR (n = 35)	R (n = 85)	P
Age, Mean ± SD	53.60 ± 10.91	53.31 ± 11.93	53.72 ± 10.54	0.855
WBC, Mean ± SD	6.23 ± 2.34	5.87 ± 1.75	6.38 ± 2.54	0.280
Neutrophils, Mean ± SD	4.98 ± 6.92	6.14 ± 12.38	4.50 ± 2.25	0.238
Lymphocytes, Mean ± SD	2.84 ± 16.96	1.18 ± 0.39	3.52 ± 20.15	0.494
RBC, Mean ± SD	4.56 ± 0.75	4.58 ± 0.66	4.55 ± 0.79	0.847
Hemoglobin, Mean ± SD	136.82 ± 21.87	138.83 ± 21.00	136.00 ± 22.29	0.522
Platelets, Mean ± SD	184.64 ± 84.79	159.46 ± 77.51	195.01 ± 85.91	0.036
AST, Mean ± SD	80.44 ± 59.60	82.03 ± 74.35	79.78 ± 52.84	0.852
ALT, Mean ± SD	54.79 ± 40.21	49.45 ± 26.14	57.00 ± 44.69	0.352
Albumin, Mean ± SD	39.14 ± 7.24	38.87 ± 5.93	39.25 ± 7.74	0.794
ALP, Mean ± SD	159.01 ± 96.09	153.37 ± 121.98	161.34 ± 83.90	0.681
GGT, Mean ± SD	252.40 ± 302.25	256.87 ± 274.62	250.56 ± 314.46	0.918
DBIL, Mean ± SD	5.44 ± 3.39	5.35 ± 3.30	5.48 ± 3.44	0.840
TBIL, Mean ± SD	18.71 ± 7.73	18.91 ± 7.96	18.63 ± 7.68	0.856
PT, Mean ± SD	13.54 ± 1.32	13.47 ± 1.11	13.57 ± 1.40	0.712
Creatinine, Mean ± SD	70.94 ± 23.33	65.68 ± 15.09	73.11 ± 25.74	0.113
Na+, Mean ± SD	138.38 ± 2.69	138.52 ± 2.48	138.33 ± 2.79	0.719
K+, Mean ± SD	4.10 ± 0.34	4.10 ± 0.35	4.10 ± 0.33	0.989
Tumor Diameter, Mean ± SD	94.78 ± 36.72	88.53 ± 41.98	97.35 ± 34.25	0.276
Gender, n (%)				1.000
Male	106 (88.33)	31 (88.57)	75 (88.24)
Female	14 (11.67)	4 (11.43)	10 (11.76)
AFP, n (%)				0.021
≤ 400 µg/L	46 (38.33)	19 (54.29)	27 (31.76)
> 400 µg/L	74 (61.67)	16 (45.71)	58 (68.24)

Table 2.

Multivariate logistic regression analysis of clinical features in the HAIC cohort patients

Variables	β	S.E	Z	P	OR (95% CI)
AFP (> 400 µg/L)	0.97	0.42	2.30	0.022	2.63 (1.15 ~ 6.02)
Platelets	0.01	0.00	2.08	0.038	1.01 (1.01 ~ 1.01)

Additionally, among 493 HCC patients who underwent surgical resection, 156 were excluded due to prior pretreatment, 109 for poor image quality, and 121 for missing postoperative pathological data. Ultimately, 107 patients were enrolled in the surgical resection cohort, with a mean age of 56 (± 10.94) years, of which 84.11% were male, and 15.89% were female (Table 3).

Table 3.

Baseline characteristics of HCC patients underwent surgical resection

Variables	Total (n = 107)
Age, Mean ± SD	56.38 ± 10.94
WBC, Mean ± SD	6.07 ± 2.29
Neutrophils, Mean ± SD	3.98 ± 2.12
Lymphocytes, Mean ± SD	2.00 ± 4.15
RBC, Mean ± SD	4.56 ± 0.55
Hemoglobin, Mean ± SD	141.75 ± 17.48
Platelets, Mean ± SD	183.52 ± 82.43
AST, Mean ± SD	65.31 ± 70.44
ALT, Mean ± SD	69.12 ± 84.52
Albumin, Mean ± SD	42.08 ± 8.01
ALP, Mean ± SD	168.73 ± 213.12
GGT, Mean ± SD	165.85 ± 202.59
DBIL, Mean ± SD	9.29 ± 29.20
TBIL, Mean ± SD	27.73 ± 44.20
Creatinine, Mean ± SD	65.67 ± 14.90
Na+, Mean ± SD	140.31 ± 2.98
K+, Mean ± SD	4.06 ± 0.60
PT, Mean ± SD	11.89 ± 1.22
Tumor Diameter, Mean ± SD	38.35 ± 11.12
Gender, n (%)
Male	90 (84.11)
Female	17 (15.89)
AFP, n (%)
≤ 400 µg/L	81 (75.70)
> 400 µg/L	26 (24.30)

Habitat imaging and radiomic features

Tumor habitats were systematically delineated through CEMRI preprocessing, tumor segmentation, feature map calculation, and clustering analysis (Fig. 3). After obtaining AN and HN maps (Fig. 3A-B), the optimal cluster number was determined as k = 3 (Figs. 3C-D) based on the elbow method and silhouette coefficient, dividing the tumors into three distinct habitats characterized by low, medium, and high enhancement patterns. The AN and HN values for the three habitats are presented as follows: Habitat 1, AN = 0.0164 ± 0.0491, HN = 0.4823 ± 0.0637; Habitat 2, AN = 0.6523 ± 0.1001, HN = 0.5356 ± 0.1321; Habitat 3, AN = 0.3939 ± 0.0787, HN = 0.4808 ± 0.1181 (Fig. 3E). Among all patients, the proportions of the Habitat 1, 2, and 3 were 21.4%, 36.4%, and 42.2%, respectively (Fig. 3F). A pronounced difference in the spatial distribution of these habitats was observed between patients who responded well and those who responded poorly to HAIC-FOLFOX therapy. Responders predominantly exhibited Habitat 3, whereas non-responders showed predominance of Habitat 1 (Fig. 4A-B). In addition, we found that the proportions of the three habitats in the AN map were 0.9%, 58.3%, and 40.8%, respectively, while those in the HN map were 20.6%, 38.9%, and 40.4%, respectively (Fig. 3G-H), revealing the significant differences in the composition of internal heterogeneity between the two phases. Subsequently, LASSO regression was employed to select the most relevant features for the final model (Figs. 5A-B): Habitat 1 voxels (habitat1), Habitat 3 voxels (habitat3), interaction ratio between liver parenchyma and habitat 3 (b03), and interaction ratio between habitat 1 and habitat 2 (1b2r).

Fig. 4.

Representative CEMRI images and habitat maps of two HCC patients. (A) The HAIC-responsive (R) patient, whose habitat map was generated from both arterial phase (AP) and hepatobiliary phase (HBP) of CEMRI, was dominated by Habitat 3. (B) The HAIC-non-responsive (NR) patient was dominated by Habitat 1

Fig. 5.

Performance evaluation and comparison of the models for predicting the therapeutic response of HAIC-FOLFOX treatment. (A-B) Habitat model feature selection using LASSO regularization. (C) Receiver operating characteristic (ROC) curves demonstrating the discriminatory performance of different predictive models. (D) Calibration curves comparing predicted versus observed probabilities for each model, the dashed diagonal line represents perfect calibration. (E) Decision curve analysis illustrating the net benefit of each model across threshold probabilities. (F) Radar chart for comparing performance indicators of different predictive models

For conventional radiomics, LASSO regression was likewise utilized to select the optimal features for model construction, resulting in the retention of the five most discriminative features, including Dependence NonUniformity Normalized (DNN), Large Dependence Emphasis (LDE), Large Dependence Low Gray Level Emphasis (LDLGLE), Gray Level NonUniformity (GLN), and Busyness (BN).

Model construction

A habitat risk score was derived from the final retained habitat features using binary logistic regression, defined by the following formula:

The clinical indicators associated with HAIC-FOLFOX treatment response are listed in Table 2. A clinical scoring model was then constructed using logistic regression with the following formula:

Based on the final retained conventional radiomics features, a model was constructed using binary logistic regression. The formula is as follows:

The clinical indicators were combined with the habitat risk score to develop a composite prediction model, as defined by the following formula:

Model performance

For predicting early HAIC-FOLFOX response, the habitat radiomics, conventional radiomics, and clinical models achieved AUCs of 0.868 (95% CI: 0.748–0.976), 0.849 (95% CI: 0.719–0.954), and 0.653 (95% CI: 0.497–0.802), respectively (Fig. 5C). Pairwise comparisons using DeLong’s test revealed that the habitat radiomics model significantly outperformed the conventional radiomics and clinical models. Furthermore, we developed a clinical-habitat combined model, achieving an AUC of 0.901 (95% CI: 0.795–0.989), which was superior to all single-modality models (Fig. 5C). Internal validation with 1,000 bootstrap resamples was performed to evaluate the stability and reliability of the model. The calibration curve demonstrated that the prediction probability of the habitat radiomics model was highly consistent with the actual observation probability (Fig. 5D). At threshold probabilities of 0.3–0.6, the decision curve showed that the combined model achieved the highest net clinical benefit, followed by the habitat radiomics model (Fig. 5E). However, within the 0.6–0.8 threshold range, the habitat radiomics model surpassed the combined model in net clinical benefit (Fig. 5E). The radar charts were utilized to provide an intuitive comparison of these models across six metrics: AUC, accuracy, sensitivity, specificity, true positive rate (TPR), and true negative rate (TNR) (Fig. 5F). The combined model consistently outperformed the other three across all dimensions, reflecting its comprehensive and superior predictive capability. The habitat radiomics model also demonstrated better overall performance than both the conventional radiomics and clinical models.

Robust performance of radiomics models across cirrhosis status

To further assess the impact of cirrhosis on model performance, we stratified the HAIC-treated cohort into the cirrhosis subgroup (R = 71, NR = 27) and the non-cirrhosis subgroup (R = 14, NR = 8). Within the cirrhosis subgroup, the combined model achieved the highest discriminatory performance, with an AUC of 0.901 (95% CI: 0.785–0.991) and optimal comprehensive metrics (Fig. 6A). The habitat radiomics model and the conventional radiomics model demonstrated comparable performance, with AUCs of 0.790 (95% CI: 0.595–0.958) and 0.786 (95% CI: 0.612–0.926), respectively, but both were inferior to the combined model (Fig. 6A). The clinical model demonstrated limited discriminatory performance (AUC: 0.534, 95% CI: 0.337–0.678) (Fig. 6A). In the non-cirrhosis subgroup, the habitat radiomics model maintained the best single-modality performance (AUC: 0.817, 95% CI: 0.500-1.000), with the combined model (AUC: 0.802, 95% CI: 0.229-1.000) and conventional radiomics model (AUC: 0.752, 95% CI: 0.291-1.000) following closely, and the clinical model exhibiting moderate improvement (AUC: 0.708, 95% CI: 0.333-1.000) but still lagging behind radiomics-based models (Fig. 6B). Calibration curves indicated that the habitat radiomics and combined models were well-calibrated in both subgroups, whereas the conventional radiomics and clinical models exhibited minor deviations (Fig. 6C-D). Decision curve analysis revealed that within the cirrhosis subgroup, the combined model provided the highest net benefit across most threshold probabilities (0.0-0.8), particularly within the clinically relevant 0.3–0.6 range, where it significantly outperformed other models (Fig. 6E). In the non-cirrhosis subgroup, the habitat radiomics and combined models demonstrated comparable net benefit at thresholds of 0.0-0.5, while the habitat radiomics model slightly exceeded the combined model at thresholds of 0.5–0.8 (Fig. 6F). Overall, the performance trends of the models were consistent between the two subgroups, indicating that the predictive value of both the habitat radiomics and combined models is independent of liver cirrhosis status, which supports their broad clinical applicability for guiding personalized HAIC-FOLFOX treatment decisions.

Fig. 6.

Performance evaluation of prediction models for early HAIC-FOLFOX response stratified by liver cirrhosis status. (A, C, E) ROC curve, calibration curve, and decision curve analysis of each model for the cirrhosis subgroup. (B, D, F) ROC curve, calibration curve, and decision curve analysis of each model for the non-cirrhosis subgroup

Habitat radiomics-derived pathological features

Subsequently, the patients in the surgical resection cohort were stratified into high-risk (n = 66) and low-risk (n = 41) groups using the habitat model (Fig. 7A). Eighteen objective pathomic features (Fig. 7B) characterizing nuclear and cytoplasmic morphology were extracted from histological whole-slide images, including Morphological features (Eccentricity, Circularity, Elongation, Major Axis Length, Minor Axis Length), Shape features (Area, Area Bbox, Perimeter, Solidity, Extent), Curvature features (CurvMean, CurvMax, CurvMin, CurvStd), and Intensity features (Intensity Mean, Intensity Std, Intensity Max, Intensity Min). Among these pathomic features, IntensityMin and CurvMean (Figs. 7D-E) were significantly associated with predicted outcomes (P < 0.05). Regarding the 10 IHC markers analyzed (AFP, Arg-1, CD34, CK19, Hepatocyte, Ki-67, P53, Glypican-3, MVI, Pathological grading) (Fig. 7C), CD34, CK19, and GPC3 (Figs. 7F-H) exhibited significant correlations with predicted outcomes (P < 0.05). Figure 8 presents representative patients from the two risk groups identified by the habitat model with distinct predicted outcomes. Specifically, patients in the low-risk group displayed significantly higher IntensityMin values and elevated CD34 expression (Fig. 8A and B), along with lower CurvMean values and reduced expression of CK19 and GPC3 (Fig. 8C and D). In contrast, the high-risk group exhibited remarkably lower IntensityMin values and CD34 expression (Fig. 8E and F), as well as significantly higher CurvMean values and increased expression of CK19 and GPC3 (Fig. 8G and H). Correlation analysis further revealed a significant negative association between GPC3 and IntensityMin (Fig. 9A), and between CD34 and CurvMean (Fig. 9B)(P < 0.05).

Fig. 7.

Pathomics and Immunohistochemical Analysis. (A) Outcome prediction for the surgical resection cohort using the habitat-based predictive model. (B) Correlation between pathological features and predictive outcomes. (C) Correlation between immunohistochemical markers and predictive outcomes. (D-E) Distribution patterns of pathological features. (F-H) Distribution patterns of immunohistochemical features

Fig. 8.

Representative pathological images of two hepatectomy patients. (A-D) HE staining and immunohistochemical staining images of pathological sections from patients predicted as low-risk by the habitat model. (E-H) HE staining and immunohistochemical staining images of pathological sections from patients predicted as high-risk by the habitat model

Fig. 9.

Correlation between pathomic and immunohistochemical markers. (A) Correlation dot plot of the pathomic feature CurvMean and the immunohistochemical marker CD34. P values were determined by Spearman correlation analysis. (B) Correlation dot plot of the pathomic feature Intensity Min and the immunohistochemical marker Glypican-3

Discussion

The therapeutic effect of HAIC-FOLFOX on HCC is significantly restricted by tumor heterogeneity [26]. Therefore, precise assessment of tumor heterogeneity and accurate prediction of treatment response are crucial for optimizing clinical decision-making. Recent studies have confirmed that habitat radiomics exhibits superior performance compared to traditional radiomics in predicting tumor treatment effects due to its unique advantages in quantifying tumor spatial heterogeneity [37–40]. In this study, we pioneered the development of a CEMRI-based habitat radiomics model for predicting early treatment response to HAIC-FOLFOX in patients with unresectable HCC. The habitat radiomics model proved robust performance, achieving an AUC of 0.868, which outperformed both the conventional radiomics (AUC = 0.849) and clinical models (AUC = 0.653) (P < 0.05, DeLong test). Moreover, integration of clinical features, the AUC of the combined model was 0.901, with decision curve analysis confirming significant net clinical benefits and highlighting the practical value of multimodal fusion. Subsequently, we further verified the clinical applicability of the habitat model based on liver cirrhosis status within the HAIC-treated cohort. Compared with the conventional radiomics model and clinical model, the combined model and habitat radiomics model demonstrated superior and more stable predictive performance. In the non-cirrhotic subgroup, the habitat and combined models retained favorable predictive ability, although their accuracy was somewhat diminished by the limited sample size. Importantly, we further established mechanistic links between imaging features and underlying pathological/molecular alterations, endowing the habitat model with interpretable biological insights.

By applying clustering analysis to CEMRI data, we identified three subregions with divergent HAIC response associations. Subsequently, the MSI matrix analysis was used to quantify the spatial heterogeneity of tumors [39, 41]. The results revealed that “habitat3” and “b03” were protective factors for an early treatment response to HAIC-FOLFOX, while “habitat1” and “1b2r” were identified as risk factors. Specifically, habitat 1 demonstrated persistently low enhancement across both arterial and hepatobiliary phases, in contrast to habitat 3, which exhibited high enhancement in both phases. This pattern was consistent with prior observations that high-perfusion regions improve arterial chemotherapy delivery [40].

Additionally, this study integrated MRI habitat features with microscopic pathological features, enhancing the interpretability of the model. The inverse correlation between CurvMean and treatment response indicates that tumors with more irregular glandular or nuclear boundaries (reflected by higher curvature values) exhibit heightened resistance to HAIC. This morphological complexity typically signifies high-grade cellular pleomorphism and aggressive tumor biology. Conversely, the positive correlation between IntensityMin and treatment response suggests that regions with higher nuclear density or deeper staining on H&E sections, potentially indicative of proliferative activity, exhibit superior responsiveness to HAIC. These well-perfused regions may facilitate more effective drug delivery through arterial infusion [42]. These findings confirm MRI habitat features are not mere statistical constructs but surrogate markers of pathological heterogeneity, bridging non-invasive imaging and invasive histopathology.

Furthermore, our IHC results unraveled molecular mechanisms of HAIC response and their association with pathomic features. CD34-positive microvessel density (MVD) correlated positively with efficacy, emphasizing the critical role of vascular supply in arterial chemotherapy delivery [43]. In contrast, the negative correlations of CK19 and GPC3 with treatment response identify a chemoresistant tumor subtype. CK19 marks hepatic progenitor cells and defines aggressive HCC subtypes characterized by epithelial-mesenchymal transition and enhanced DNA repair capacity [44]. Similarly, GPC3, through its involvement in Wnt/β-catenin and YAP signaling pathways, promotes tumor proliferation and survival while concurrently conferring treatment resistance [45, 46]. Importantly, IHC markers also correlated with pathomic features. An elevated CD34-MVD level was associated with a lower CurvMean value, suggesting that sufficient vascularization helps maintain tissue architectural integrity. Conversely, increased GPC3 expression inversely correlated with IntensityMin values, which may reflect the nuclear atypia and cellular crowding characteristic of highly malignant phenotypes.

These observations form a unified framework where MRI habitat features indirectly reflect vascular, morphological, and molecular characteristics shaping treatment response.

Overall, this study for the first time applied the CEMRI-based habitat radiomics model to predict the early efficacy of HAIC-FOLFOX therapy in patients with unresectable HCC, breaking through the limitation of traditional radiomics that regards the tumor as a homogeneous entity. By quantifying the spatial distribution and interaction patterns of distinct enhancement subregions within tumors, our model enables precise characterization of tumor spatial heterogeneity. Subsequently, the multi-dimensional data of pathology and immunohistochemistry were innovatively integrated, and the biological basis of habitat imaging features was systematically explained from the microscopic morphology and molecular expression level, and the cross-omics association of “imaging-pathology-molecular” was established, which provided a mechanistic interpretability for the model. Furthermore, the integrated clinical-habitat model exhibits superior predictive performance (AUC = 0.901) and possesses significant translational potential, providing a non-invasive and highly efficient tool for personalized HAIC efficacy prediction in HCC.

However, several limitations of this study warrant acknowledgment. First, the retrospective design may introduce potential selection biases and information biases, thereby limiting the generalizability of research results and the robustness of causal inferences. Second, the modest sample size and lack of external validation necessitate cautious interpretation of the results. Future prospective multicenter studies with larger cohorts are imperative to validate our findings. Third, the exclusive reliance on T1-weighted imaging for model construction may neglect complementary prognostic information from other MRI sequences, such as T2-weighted or diffusion-weighted imaging [47]. Finally, although our pathological correlation provides valuable mechanistic insights, the use of surgical specimens from a different patient cohort, rather than the HAIC-treated population, imposes inherent constraints on the interpretability of our mechanistic findings.

Conclusion

The CEMRI-based habitat radiomics model demonstrates robust predictive capability for early HAIC-FOLFOX treatment response and reveals significant associations with underlying pathomorphological and molecular characteristics. This framework establishes a promising non-invasive modality for informing personalized therapeutic strategies in HCC patients.

Abbreviations

HCC

Hepatocellular Carcinoma

CEMRI

Contrast-Enhanced Magnetic Resonance Imaging

HAIC-FOLFOX

Hepatic Arterial Infusion Chemotherapy with Fluorouracil, Leucovorin, and Oxaliplatin

ALT

Alanine Aminotransferase

AST

Aspartate Aminotransferase

LASSO

Least Absolute Shrinkage and Selection Operator

AUC

Area Under the Curve

ROC

Receiver Operating Characteristic

TPR

True Positive Rate

TNR

True Negative Rate

CI

Confidence Interval

ROI

Region of Interest

ECOG

Eastern Cooperative Oncology Group

TACE

Transcatheter Arterial Chemoembolization

OR

Objective Response

ORR

Objective Response Rate

R

Responders

NR

Non-Responders

PVTT

Portal Vein Tumor Thrombus

CT

Computed Tomography

AASLD

American Association for the Study of Liver Disease

ECOG

Eastern Cooperative Oncology Group

TACE

Transarterial Chemoembolization

mRECIST

Modified Response Evaluation Criteria in Solid Tumors

DCR

Disease Control Rate

GPC3

Glypican-3

DSA

Digital subtraction angiography

CR

Complete Response

PR

Partial Response

SD

Stable Disease

PD

Progressive Disease

T1WI

T1-Weighted Imaging

AP

Arterial Phase

VP

Venous Phase

DP

Delayed Phase

HBP

Hepatobiliary Phase

AN

Arterial Phase Enhancement

HN

Hepatobiliary Phase Enhancement

ICCs

Intraclass Correlation Coefficients

GLCM

Gray Level Co-occurrence Matrix

GLDM

Gray Level Dependence Matrix

GLRLM

Gray Level Run Length Matrix

GLSZM

Gray Level Size Zone Matrix

NGTDM

Neighboring Gray Tone Difference Matrix

H&E

Hematoxylin and eosin

IHC

immunohistochemical

AFP

Alpha-Fetoprotein

DNN

Dependence Non Uniformity Normalized

LDE

Large Dependence Emphasis

LDLGLE

Large Dependence Low Gray Level Emphasis

GLN

Gray Level NonUniformity

BN

Busyness

Author contributions

Guanhui Li: Conceptualization, Methodology, Data curation, Formal analysis, Visualization, Writing-original draft, Writing-review & editing. Xuefei Zhao: Conceptualization, Formal analysis, Visualization, Validation, Writing-original draft, Writing-review & editing. Jie Long: Writing-original draft. Yi Li: Writing-original draft. Zhexuan Ye: Investigation. Lili Rao: Investigation. Yuqin He: Investigation. Shujie Lai: Investigation. Yuxin Tang: Investigation. Hao Zhong: Investigation. Chao Li: Investigation. Jie Li: Investigation. Changxu Cai: Investigation. Hao Wu: Investigation. Xiang Lan: Investigation. Nan You: Investigation. Jun Wang: Supervision, Writing-review & editing. Liangzhi Wen: Conceptualization, Project administration, Supervision, Funding acquisition, Writing-review & editing.

Funding

This work was supported by the National Natural Science Foundation of China (NSFC: 82273484), the National Natural Science Foundation of China (NSFC: 82473105), and the project of Chongqing young and middle-aged medical talents.

Data availability

Data will be made available on request.

Declarations

Ethics approval and consent to participate

Following the Declaration of Helsinki, the study received ethical approval from the Ethics Committee of Daping Hospital of Army Medical University (Approval No. 2025(313)), the Ethics Committee of the First Affiliated Hospital of Chongqing Medical University, and the Ethics Committee of the Second Affiliated Hospital of Army Medical University. Informed consent was obtained in advance from all patients participating in the study, permitting the use of their medical records for this research and the publication of the survey results in academic journals.

Competing interests

The authors declare no competing interests.