Can incorporating preoperative arterial enhancement ratio map into habitat analysis predict TACE refractoriness in unresectable hepatocellular carcinoma? Authors: Yan Xiao, Wenlong Song, Huan Liu, Qin Yang, Jun Zhou, Dajing Guo Published: 2026 ABSTRACT AbstractPurposeTo investigate the value of applying preoperative magnetic resonance imaging (MRI)-based arterial enhancement ratio (AER) map to habitat analysis for predicting transarterial chemoembolization (TACE) refractoriness in patients with unresectable hepatocellular carcinoma (HCC).Methods176 patients with unresectable HCC who underwent preoperative contrast-enhanced MRI and received consecutive TACE treatments were randomly allocated to a training cohort (n = 124) and a validation cohort (n = 52). A dual-mode encoding strategy (mode 1: based on T1-pre; mode 2: based on T1-pre with the AER map) was employed to define tumor habitats. Signal intensity, entropy, and volume fraction were quantified for each habitat. Combined with laboratory findings and imaging features, logistic regression analyses were performed to identify independent risk factors for TACE refractoriness and a predictive model was constructed. The performance of the model was evaluated using the area under the receiver operating characteristic curve (AUC).ResultsEach tumor lesion was segmented into 5 habitats using mode 1 and 3 habitats using mode 2. Regression analysis identified Barcelona Clinic Liver Cancer stage (OR = 2.099, P = 0.027), blood products in mass (OR = 11.063, P = 0.028), and AER entropy value of habitat 2 derived from mode 2 (OR = 4.586, P = 0.033) as independent factors for predicting TACE refractoriness. AUC value of the nomogram on validation cohort was 0.721 (95% CI = 0.577–0.866). Calibration curve demonstrated favorable clinical applicability of the model.ConclusionIncorporation of the AER map into habitat analysis enables preoperative prediction of TACE refractoriness risk in patients with unresectable HCC.Clinical trial numberNot applicable.Supplementary InformationThe online version contains supplementary material available at 10.1186/s12880-026-02371-7.Keywords: Hepatocellular carcinoma, Transarterial chemoembolization, Refractoriness, Habitat, Magnetic resonance imaging Introduction Liver cancer ranks as the third leading cause of cancer-related mortality worldwide, with hepatocellular carcinoma (HCC) being the most prevalent pathological subtype [1]. For patients with unresectable HCC, transarterial chemoembolization (TACE) serves as a standard locoregional therapy, and multiple TACE sessions are usually required to achieve satisfactory tumor control [2]. Unfortunately, owing to tumor heterogeneity, some patients may experience disease progression or poor response during repeated procedures [3], a condition defined as TACE refractoriness. Accumulating evidence suggests that patients who develop TACE refractoriness should shift to systemic therapy in a timely manner to prevent further liver function deterioration and improve survival [4–6]. Consequently, preoperative identification of patients who are more prone to TACE refractoriness is crucial for optimizing treatment strategies. Currently, preliminary studies have explored the predictive value of conventional imaging features (such as maximum tumor diameter, number, and distribution) [7–9] or radiomics for TACE refractoriness [10, 11]. Although conventional imaging features are easily accessible, they only capture the macroscopic structure of tumors and overlook their inherent nature as heterogeneous entities. Radiomics, by extracting high-throughput features, partially overcomes the subjectivity of conventional indicators; however, most of these features lack definite biological explanations and fail to reflect the distributional differences of intratumoral heterogeneity. Therefore, the aforementioned “global” analytical approaches are inadequate for revealing the high heterogeneity and diversity of intratumoral cellular characteristics, which is a critical factor underlying the complexity of treatment response in HCC [12]. The emergence of habitat analysis offers a new strategy to address such limitations: by analyzing differences in imaging characteristics, it segments tumors into multiple subregions with distinct metabolic or functional profiles, thereby visually revealing the spatial heterogeneity related to tumor function. Habitat analysis has demonstrated considerable potential in the field of HCC, with multiple studies [13–16] successfully validating its value in predicting microvascular invasion, postoperative recurrence risk, and tumor composition. However, despite its excellent performance in diagnostic and prognostic assessment, application in TACE refractoriness remains unreported. The characteristic dominance of hepatic arterial supply in HCC provides a critical theoretical basis for TACE and is closely associated with postoperative outcomes [17]. As a quantitative imaging biomarker, the arterial enhancement ratio (AER) has been developed to characterize such arterial dominance with greater precision. Derived from contrast-enhanced magnetic resonance imaging (MRI), the AER effectively quantifies intratumoral perfusion differences and reflects tumor angiogenesis status [18]. On the other hand, previous study [19] has shown that textural features of HCC—such as entropy—undergo significant changes before and after TACE treatment, suggesting a potentially meaningful correlation with therapeutic response. Given the above findings, we hypothesize that utilizing AER in combination with entropy to perform habitat segmentation within tumor may effectively reveal spatial heterogeneity in vascular distribution, thereby establishing a link with TACE refractoriness. The proposed method is expected to provide a novel, noninvasive imaging biomarker for preoperative prediction of TACE refractoriness in patients with unresectable HCC. In this study, we aimed to explore whether heterogeneity analysis based on the preoperative AER map can characterize intratumoral vascular differences by comparing dual-mode habitat encoding, thereby predicting the occurrence of TACE refractoriness in patients with unresectable HCC. Furthermore, by constructing a nomogram, we seek to identify those patients who would truly benefit from repeated TACE treatment, which holds significant implications for guiding personalized therapeutic strategies. Materials and methods Patients The study complied with the Declaration of Helsinki and was approved by the Ethics Committee of the Second Affiliated Hospital of Chongqing Medical University (ID: 2021 − 240). Given the retrospective nature of the research, informed consent was waived. Figure 1 presents the patient enrollment workflow. A total of 1362 patients diagnosed with unresectable HCC by a multidisciplinary liver cancer team and treated with TACE between November 2018 and May 2024 were retrospectively collected, all of whom underwent preoperative contrast-enhanced MRI. Exclusion criteria were as follows: (1) incomplete clinical or imaging data (n = 387); (2) coexistence of other malignancies or severe comorbidities (n = 55); (3) Child-Pugh grade C (n = 77) or Barcelona Clinic Liver Cancer (BCLC) stage C (n = 102); (4) severe adverse reactions after TACE (n = 1); (5) additional antitumor therapies were received (n = 389) or with single TACE (n = 143); (6) lost follow up (n = 32). Ultimately, 176 patients were enrolled and randomly allocated to the training cohort (n = 124) and validation cohort (n = 52) at a ratio of 7: 3 using stratified sampling. Flowchart of the enrollment process. HCC = hepatocellular carcinoma, TACE = transarterial chemoembolization, MRI = magnetic resonance imaging, BCLC = Barcelona Clinic Liver Cancer Demographic and clinical data were extracted from electronic medical record system, including sex, age, liver disease etiology, cirrhosis status, alpha-fetoprotein (AFP), alkaline phosphatase (ALP), total bilirubin (TB), direct bilirubin (DB), indirect bilirubin (IB), albumin (ALB), lymphocyte count, platelet count (PLT), Child-Pugh grade, and BCLC stage. TACE procedure All TACE procedures were performed by two interventional radiologists with 10 and 20 years of interventional experience, respectively. Of these procedures, 103 patients were treated under the guidance of digital subtraction angiography system 1 (Infinix INFX-8000 V, Toshiba, Japan), while the remaining 73 patients were treated using the angiography system 2 (Infinix INFX-9000 C, Toshiba, Japan). Vascular access was established via percutaneous puncture of the femoral artery using the Seldinger technique. A 5 F catheter (RH catheter; Terumo, Tokyo, Japan) was selectively cannulated into the common hepatic artery and superior mesenteric artery for angiographic evaluation of the tumor-feeding arteries. Upon confirmation, a 2.7 F microcatheter (Progreat Microcatheter System, Terumo) was exchanged and superselectively cannulated as far as possible into the feeding arteries, following which an emulsion consisting of cisplatin (10–50 mg), pirarubicin hydrochloride (10–40 mg), and lipiodol (10–25 mL) was slowly injected. The dosage of this emulsion was determined based on the tumor characteristics and the patient’s liver function. Subsequently, the feeding arteries were further embolized using gelatin sponge particles (1–2 mm in diameter) or polyvinyl alcohol (PVA, size 150–350 μm, Alicon, Hangzhou, China). The procedural endpoint was defined as dense lipiodol accumulation within the tumor coupled with visualization of small perilesional portal vein branches. TACE treatment followed the “on-demand” principle, with decisions regarding repeat TACE sessions made according to postprocedural tumor response, liver function, and performance status, among others. Evaluation and follow-up of TACE refractoriness The study adopted 2021 revised expert consensus criteria established by Japan Society of Hepatology (JSH) [4], with whether TACE refractoriness occurs as evaluation endpoint. According to the criteria, TACE refractoriness was defined by meeting any of the following conditions: (1) persistent viable tumor volume exceeding 50% in target lesions after two or more consecutive TACE procedures; (2) increased number of lesions; (3) failure to achieve expected reduction in tumor markers; or (4) development of vascular invasion or extrahepatic metastasis. All cases were evaluated by enhanced imaging examination and tumor marker conducted within 1–3 months post-procedure, with independent and blinded interpretations by two senior abdominal radiologists, followed by consensus review. The follow-up period ended on March 1, 2025, at which point all patients were assessed for the occurrence of TACE refractoriness. MRI data acquisition The abdominal MRI examination sequences include in-phase and out-of-phase T1-weighted imaging (T1WI), fat-suppressed T2-weighted imaging (T2WI), and multiphase enhanced T1WI. Among 176 patients included, 103 patients underwent examinations using Scanner 1 (1.5T, HDxt, GE Medical Systems, Fairfield, OH, USA), and the remaining 73 patients had their data collected by Scanner 2 (1.5T, Magnetom Avanto, Siemens, Erlangen, Germany). The multiphase enhanced T1WI was performed by Gadodiamide (Omniscan®; GE Healthcare), at a rate of 2.0 mL/s and a dose of 0.2 mL/kg. After injection of contrast agent, arterial phase (AP), portal venous phase (PVP), and delayed phase (DP) scans were implemented at 25–30 s, 60–70 s, and 180 s, respectively. Details of MRI parameters are shown in appendix S1. Assessment of MRI features All MRI images were independently evaluated by two radiologists (with 10 and 15 years of abdominal imaging analysis experience, respectively) using double-blinded method. In cases of disagreement, final decision was made by a third senior radiologist with over 20 years of experience. According to 2018 version of Liver Imaging Reporting and Data System (LI-RADS) criteria [20], following imaging features were assessed in the largest lesion of each patient: (a) nonrim arterial phase hyperenhancement (APHE); (b) nonperipheral washout; (c) enhancing capsule; (d) nodule-in-nodule architecture; (e) mosaic architecture; (f) blood products in mass; (g) fat in mass. Additionally, other tumor-related prognostic features were evaluated [21–29]: nonsmooth tumor margin, hypovascular nodule, PVP peritumoral hypoenhancement, intratumoral artery, APHE proportion ≥ 50%, and initial treatment response. Appendix S2 details imaging feature definitions. Habitat analysis Image pre-processing Two sequences of pre-contrast T1WI (T1-pre) and post-contrast T1WI at AP (T1-AP) were utilized for further analysis. To eliminate misalignment artifacts caused by respiratory motion and other factors, we first employed the Elastix open-source toolkit within the 3D Slicer platform to register the T1-AP images (moving images) to the T1-pre images (fixed reference images), achieving spatial alignment between the two sequences. Subsequently, the AER map was computationally derived through image subtraction, mathematically defined as follows: where SI is the abbreviation of signal intensity. Tumor segmentation Volumes of interest (VOIs) for all patients were manually delineated by radiologist 1 (with 5 years of abdominal imaging experience) using ITK-SNAP software. On T1-pre images, with reference to T1-AP images, the boundaries of target lesions were outlined slice-by-layer without excluding cystic or necrotic areas. The segmentation results were reviewed by a senior radiologist with 15 years of abdominal imaging experience. For 30 randomly selected patients, MR images were independently delineated again by radiologist 2 (with 10 years of abdominal imaging experience) to assess inter-observer reproducibility using the intraclass correlation coefficient (ICC), with ICC > 0.8 indicating good agreement. Thereafter, the VOIs delineated on T1-pre images were automatically transferred to the AER maps for further analysis. Habitat encoding To identify intratumoral spatial heterogeneity, we defined distinct tumor subregions termed “habitats” based on signal intensity (SI) and local entropy from T1-pre images and AER maps. The local entropy was computed using a 9 × 9 moving window on each MR image slice to quantify regional textural complexity, with higher entropy values indicating greater heterogeneity and structural irregularity. The entropy calculation formula is as follows: And the SI (i.e., voxel value) of each voxel in T1-pre images and AER maps reflects the enhancement characteristics of HCC. Subsequently, original voxel values were fused with their corresponding local entropy maps by superimposing pixel values from the original MR images and MR entropy maps to generate fused feature maps. During habitat identification, the K-means clustering algorithm was applied to simultaneously group all voxels across the entire patient cohort (including the training and validation cohort), using squared Euclidean distance between voxel intensities as the metric. Each resulting cluster category was mapped back to the original image space to form visualized “spatial habitats.” The optimal number of clusters was determined at the population level using the Calinski–Harabasz (CH) index as the evaluation criterion. After habitat definition, feature extraction and subsequent analyses were performed independently under strict sample partitioning, with the validation cohort not involved in any parameter optimization. Habitat quantification Following habitat encoding, quantitative values for each habitat were calculated on their respective parameter maps: for mode 1, these included mean SI and mean entropy value on T1-pre images, as well as volume fraction of each habitat. For mode 2, the parameters comprised mean SI and mean entropy value derived from both T1-pre images and AER maps, in addition to volume fraction. Figure 2 presents a brief flowchart of the habitat analysis process. A brief flowchart of habitat analysis Statistical analysis Statistical analyses were performed using SPSS 27.0 and R 4.2.1 software. Continuous variables were compared using Student’s t-test or Mann-Whitney U test, while categorical variables were analyzed using chi-square test. Independent predictors of TACE refractoriness were determined via univariate and multivariate binary logistic regression analyses. Based on significant variables, a predictive model was constructed using the logistic regression algorithm. The nomogram was developed to visualize predictive model and calculate risk scores for individual patients. Predictive performance of the model was evaluated by the area under receiver operating characteristic curve (AUC) in both training and validation cohorts, while consistency was evaluated using calibration curves and the Hosmer-Lemeshow test (P > 0.05 indicates good fit). A two-tailed P-value < 0.05 was considered statistically significant. Results Patients characteristics Table 1 summarizes the clinical data and imaging features of patients in the training cohort (n = 124) and validation cohort (n = 52). The mean age of all patients was 55 years, the vast majority were male (83.5%), and a total of 160 patients (90.9%) were positive for hepatitis B virus. Comparative analysis showed that all variables were well-balanced between the two cohorts. The study cohort comprised 104 patients with TACE refractoriness and 72 without. The prevalence of TACE refractoriness was comparable between the training cohort (58.9%) and the validation cohort (59.6%), showing no statistically significant difference (P = 0.927). ICC analysis demonstrated excellent reliability of VOI delineation between the two radiologists, with a mean value of 0.912. Table 1Comparison of clinical data and imaging features between the training and validation cohortsCharacteristicsOverall (n = 176)Training cohort (n = 124)Validation cohort (n = 52) P Sex Male147104 (83.88%)43 (82.70%)0.848 Female2920 (16.12%)9 (17.30%)Age (y)55 ± 10.5656.31 ± 10.5254.17 ± 10.580.220Etiology Hepatitis B virus160114 (91.93%)46 (88.46%)0.465 Other1610 (8.07%)6 (11.54%)Cirrhosis Yes148105 (84.68%)43 (82.70%)0.743 No2819 (15.32%)9 (17.30%)AFP(ug/L)45.44 (7.79, 262.48)45.50 (7.17, 202.73)42.60 (8.42, 363.85)0.548ALP(U/L)96.00 (73.25, 120.75)96.50 (72.00, 124.00)95.50 (76.75, 104.00)0.420TB (µmol/L)17.95 (12.96, 27.98)18.65 (15.43, 32.80)15.95 (11.65, 23.83)0.059DB(µmol/L)7.10 (4.53, 11.88)7.30 (5.43, 12.88)6.90 (4.33, 10.25)0.062IB(µmol/L)10.45 (7.40, 15.10)10.95 (8.05, 15.30)9.60 (6.75, 14.40)0.076ALB (g/L)39.13 ± 5.8238.77 ± 5.9940.00 ± 5.360.200Lymphocyte1.01 (0.69, 1.58)1.05 (0.68, 1.58)0.91 (0.71, 1.43)0.372PLT84.00 (55.25, 129.00)84.00 (53.25, 124.50)84.00 (60.00, 137.50)0.520BCLC Stage 03726 (20.97%)11 (21.15%)0.993 A10675 (60.48%)31 (59.62%) B3323 (18.55%)10 (19.23%)Child-Pugh Grade A13392 (74.19%)41 (78.85%)0.512 B4332 (25.81%)11 (21.15%)No.of tumor 111479 (63.71%)35 (67.30%)0.648 ≥ 26245 (36.29%)17 (32.70%)Tumor distribution Unilobar128109 (87.91%)40 (76.92%)0.065 Bilobar2815 (12.09%)12 (22.08%) Largest tumor size(mm)28.25 (18.53, 41.28)28.45 (18.20, 39.93)27.95 (18.68, 51.38)0.622Radiologic characteristicsNonrim APHE Yes12788 (70.97%)39 (75.00%)0.586 No4936 (29.03%)13 (25.00%)Nonperipheral washout Yes13694 (75.81%)42 (80.77%)0.474 No4030 (24.19%)10 (19.23%)Enhancing capsule Yes4934 (27.41%)15 (28.85%)0.847 No12790 (72.59%)37 (71.15%)Nodule in nodule architecture Yes137 (5.65%)7 (13.46%)0.080 No163117 (94.35%)45 (86.54%)Mosaic architecture Yes6345 (36.29%)18 (34.62%)0.833 No11379 (63.71%)34 (65.38%)Blood products in mass Yes1612 (9.68%)4 (7.69%)0.676 No160112 (90.32%)48 (92.31%)Fat in mass Yes3924 (19.35%)15 (28.85%)0.167 No137100 (80.65%)37 (71.15%)Nonsmooth tumor margin Yes7448 (38.71%)26 (50.00%)0.166 No10276 (61.29%)26 (50.00%)Hypovascular nodule Yes52 (1.62%)3 (5.77%)0.130 No171122 (98.38%)49 (94.23%)PVP peritumoral hypoenhancement Yes74 (3.23%)3 (5.77%)0.431 No169120 (96.77%)49 (94.23%)Intratumoral artery Yes2514 (11.30%)11 (21.15%)0.087 No151110 (88.70%)41 (78.85%)APHE proportion ≥ 50% Yes13594 (75.80%)41 (78.85%)0.663 No4130 (24.20%)11 (21.15%)Initial treatment responseReponse group139101 (81.45%)38 (73.07%)0.213Non-reponse group3723 (18.55%)14 (26.93%)TACE refractoriness Yes10473 (58.87%)31 (59.62%)0.927 No7251 (41.13%)21 (40.38%)Continuous variables are expressed as means (± standard deviations) or medians (interquartile ranges), and categorical variables are expressed as numbers (percentages). AFP = alpha-fetoprotein, ALP = alkaline phosphatase, TB = total bilirubin, DB = direct bilirubin, IB = indirect bilirubin, ALB = albumin, PLT = platelet, BCLC = Barcelona Clinic Liver Cancer, APHE = arterial phase hyperenhancement, PVP = portal venous phase, TACE = transarterial chemoembolization Comparison of clinical data and imaging features between the training and validation cohorts Continuous variables are expressed as means (± standard deviations) or medians (interquartile ranges), and categorical variables are expressed as numbers (percentages). AFP = alpha-fetoprotein, ALP = alkaline phosphatase, TB = total bilirubin, DB = direct bilirubin, IB = indirect bilirubin, ALB = albumin, PLT = platelet, BCLC = Barcelona Clinic Liver Cancer, APHE = arterial phase hyperenhancement, PVP = portal venous phase, TACE = transarterial chemoembolization Habitat mapping As shown in Fig. 3A-B, mode 1 and mode 2 reached their peak CH index values at cluster numbers (K) of 5 and 3, respectively. Accordingly, each lesion was encoded into 5 habitats using mode 1 and 3 habitats using mode 2. The parameter characteristics of each habitat are presented in Fig. 3C-D. Under the decoding of mode 2, the SI and entropy values for each habitat were as follows: habitat 1: T1-SI = 0.38 (0.20, 0.56) × 10³, T1-entropy = 0.46 (0, 5.09), AER-SI = 1.01 (0.69, 1.42), AER-entropy = 0.81 (0.19, 0.98); habitat 2: T1-SI = 0.47 (0.23, 0.68) × 10³, T1-entropy = 0.02 (0, 4.93), AER-SI = 0.51 (0.35, 0.79), AER-entropy = 0.22 (0.01, 0.68); habitat 3: T1-SI = 0.38 (0.20, 0.56) × 10³, T1-entropy = 0.34 (0, 5.06), AER-SI = 0.59 (0.37, 0.85), AER-entropy = 0.25 (0.01, 0.64). Habitat characteristics. Calinski-Harabasz index under different number of clusters for mode 1 (A) and mode 2 (B). Distribution maps of habitat parameter characteristics for mode 1 (C) and mode 2 (D) Risk factors for TACE refractoriness Univariate logistic regression analysis revealed that: BCLC stage (OR = 2.400, 95% CI = 1.287–4.476, P = 0.006), blood products in mass (OR = 8.871, 95% CI = 1.107–71.060, P = 0.040), nonsmooth tumor margin (OR = 2.304, 95% CI = 1.069–4.966, P = 0.033), as well as habitat 2 T1-entropy (OR = 0.847, 95% CI = 0.729–0.985, P = 0.031) and AER-entropy (OR = 5.804, 95% CI = 1.741–19.353, P = 0.004) derived from mode 2 were significantly associated with TACE refractoriness (Tables 2 and 3). None of the habitat parameters derived from mode 1 (based on T1-pre) showed statistical significance. Subsequent multivariate analysis identified three independent risk factors: BCLC stage (OR = 2.099, 95% CI = 1.089–4.046, P = 0.027), blood products in mass (OR = 11.063, 95% CI = 1.304–93.880, P = 0.028), and habitat 2 AER-entropy derived from mode 2 (OR = 4.586, 95% CI = 1.127–18.669, P = 0.033) (Table 4). Representative imaging findings from a typical patient with TACE refractoriness are displayed in Fig. 4. Table 2Univariate logistic regression of clinical data and imaging features for TACE refractoriness in the training cohortVariableOdds ratio (95% CI)P valueSex (male/female)0.651 (0.249 ~ 1.701)0.381Age (yr)0.999 (0.965 ~ 1.033)0.945Etiology (HBV/other)1.697 (0.417 ~ 6.900)0.460Cirrhosis (absent/present)1.049 (0.390 ~ 2.823)0.925AFP (ug/L)1.000 (0.999 ~ 1.001)0.686ALP (U/L)1.002 (0.994 ~ 1.010)0.646TB (µmol/L)0.988 (0.965 ~ 1.012)0.342DB (µmol/L)0.990 (0.952 ~ 1.030)0.629IB (µmol/L)0.974 (0.940 ~ 1.010)0.153ALB (g/L)1.026 (0.965 ~ 1.089)0.412Lymphocyte1.175 (0.723 ~ 1.911)0.516PLT0.999 (0.993 ~ 1.005)0.759BCLC Stage (0/A/B)2.400 (1.287 ~ 4.476)0.006*Child-Pugh Grade (A/B)0.865 (0.384 ~ 1.951)0.727No.of tumor1.674 (0.782 ~ 3.586)0.185Tumor distribution (Unilobar/bilobar)0.685 (0.219 ~ 2.139)0.515Largest tumor size1.015 (0.995 ~ 1.036)0.150Nonrim APHE (absent/present)0.628 (0.279 ~ 1.413)0.261Nonperipheral washout (absent/present)1.927 (0.84 ~ 4.419)0.122Enhancing capsule (absent/present)1.402 (0.619 ~ 3.176)0.418Nodule in nodule architecture (absent/present)0.686 (0.133 ~ 3.542)0.652Mosaic architecture (absent/present)1.442 (0.678 ~ 3.067)0.342Blood products in mass (absent/present)8.871 (1.107 ~ 71.060)0.040*Fat in mass (absent/present)2.455 (0.899 ~ 6.702)0.080Nonsmooth tumor margin (absent/present)2.304 (1.069 ~ 4.966)0.033*Hypovascular nodule (absent/present)2.840 (0.311~0.25.949)0.355PVP peritumoral hypoenhancement (absent/present)0.690 (0.094 ~ 5.066)0.715Intratumoral artery (absent/present)1.865 (0.551 ~ 6.314)0.316APHE proportion ≥ 50% (absent/present)0.94 (0.407 ~ 2.173)0.885Initial treatment response (reponse group/non-RG)1.764 (0.668 ~ 4.661)0.252Data in parentheses are 95% CIs. TACE = transarterial chemoembolization, AFP = alpha-fetoprotein, ALP = alkaline phosphatase, TB = total bilirubin, DB = direct bilirubin, IB = indirect bilirubin, ALB = albumin, PLT = platelet, BCLC = Barcelona Clinic Liver Cancer, PVP = portal venous phase, APHE = arterial phase hyperenhancement Univariate logistic regression of clinical data and imaging features for TACE refractoriness in the training cohort Data in parentheses are 95% CIs. TACE = transarterial chemoembolization, AFP = alpha-fetoprotein, ALP = alkaline phosphatase, TB = total bilirubin, DB = direct bilirubin, IB = indirect bilirubin, ALB = albumin, PLT = platelet, BCLC = Barcelona Clinic Liver Cancer, PVP = portal venous phase, APHE = arterial phase hyperenhancement Table 3Univariate logistic regression of habitat parameters under different habitat strategies for TACE refractoriness in the training cohortVariableOdds ratio(95% CI)P valueMode1Habitat 1 SI1.001 (1.000 ~ 1.003)0.139Habitat 1 entopy0.867 (0.747 ~ 1.006)0.060Habitat 2 SI1.001 (1.000 ~ 1.003)0.164Habitat 2 entopy0.873 (0.754 ~ 1.011)0.070Habitat 3 SI1.001 (1.000 ~ 1.002)0.093Habitat 3 entopy0.872 (0.752 ~ 1.010)0.068Habitat 4 SI1.001 (0.999 ~ 1.003)0.492Habitat 4 entopy0.924 (0.797 ~ 1.07)0.291Habitat 5 SI1.001 (0.999 ~ 1.003)0.454Habitat 5 entopy1.423 (0.519 ~ 3.899)0.493Habitat 1 ratio389.837 (0.079 ~ 1912641.247)0.169Habitat 2 ratio0.429 (0.002 ~ 81.372)0.752Habitat 3 ratio0.009 (0 ~ 379835.515)0.599Habitat 4 ratio0.378 (0.003 ~ 56.420)0.703Habitat 5 ratio00.416Mode2Habitat 1 T1-SI1.001 (1.000 ~ 1.003)0.142Habitat 1 T1-entopy0.878 (0.758 ~ 1.017)0.083Habitat 1 AER-SI1.512 (0.817 ~ 2.801)0.188Habitat 1 AER-entopy1.778 (0.840 ~ 3.762)0.132Habitat 2 T1-SI1.001 (1.000 ~ 1.002)0.103Habitat 2 T1-entopy0.847 (0.729 ~ 0.985)0.031*Habitat 2 AER-SI1.665 (0.623 ~ 4.45)0.309Habitat 2 AER-entopy5.804 (1.741 ~ 19.353)0.004*Habitat 3 T1-SI1.001 (0.999 ~ 1.003)0.187Habitat 3 T1-entopy0.884 (0.763 ~ 1.023)0.099Habitat 3 AER-SI1.519 (0.671 ~ 3.439)0.316Habitat 3 AER-entopy2.545 (0.889 ~ 7.286)0.082Habitat 1 ratio5.138 (0.017 ~ 1523.708)0.573Habitat 2 ratio0.244 (0.001 ~ 483.541)0.716Habitat 3 ratio0.474 (0.002 ~ 103.085)0.786Data in parentheses are 95% CIs. TACE = transarterial chemoembolization, SI = signal intensity, AER = arterial enhancement ratio Univariate logistic regression of habitat parameters under different habitat strategies for TACE refractoriness in the training cohort Data in parentheses are 95% CIs. TACE = transarterial chemoembolization, SI = signal intensity, AER = arterial enhancement ratio Table 4Multivariate logistic regression of the predictors for TACE refractoriness in the training cohortPredictorβ (Coefficient)Odds ratio95% CIP valueBCLC stage0.7422.0991.089–4.0460.027*Blood products in mass2.40411.0631.304–93.8800.028*Nonsmooth tumor margin0.6952.0040.853–4.7060.111Habitat 2 T1-entopy-0.0670.9350.774–1.1300.448Habitat 2 AER-entopy1.5234.5861.127–18.6690.033*TACE = transarterial chemoembolization, BCLC = Barcelona Clinic Liver Cancer, AER = arterial enhancement ratio Multivariate logistic regression of the predictors for TACE refractoriness in the training cohort TACE = transarterial chemoembolization, BCLC = Barcelona Clinic Liver Cancer, AER = arterial enhancement ratio Representative images of an unresectable HCC patient who developed TACE refractoriness. (A) A mass in the right lobe, complicated by intratumoral hemorrhage (green arrow). (B) Three habitats segmented via mode 2 are presented using T1-AP image, where habitat 2 (red area) exhibited an AER entropy value of 0.64. (C–D) Follow-up imaging at 2 months after the second TACE session revealed portal vein tumor thrombosis (blue arrow) and pulmonary metastasis (red arrow). HCC = hepatocellular carcinoma, TACE = transarterial chemoembolization, AP = arterial phase, AER = arterial enhancement ratio Construction and evaluation of the predictive model Based on the three independent predictors (BCLC stage, blood products in mass, and habitat 2 AER-entropy) identified by logistic regression analysis, we developed and validated a predictive model for TACE refractoriness. As shown in Fig. 5A-B, the model demonstrated an AUC of 0.793 (95% CI: 0.710–0.876) in the training cohort and 0.721 (95% CI: 0.577–0.866) in the validation cohort, indicating its potential value in predicting TACE refractoriness in unresectable HCC patients. Detailed evaluation metrics of the predictive model are provided in appendix S3. Calibration curves showed good agreement between model predictions and actual observations (Fig. 5C-D). A nomogram (Fig. 6) was constructed to visualize the predictive model and calculate a total risk score from preoperative clinical and imaging features, allowing for the quantitative evaluation of individual TACE failure risk. Construction and validation of the model for predicting TACE refractoriness in unresectable HCC. ROC curves and the AUC values in the training cohort (A) and the validation cohort (B). Calibration curves for the predictive model in the training cohort (C) and validation cohort (D). TACE = transarterial chemoembolization, HCC = hepatocellular carcinoma, ROC = receiver operating characteristic, AUC = area under the ROC curve Nomogram incorporating AER map-derived habitat analysis for predicting TACE refractoriness in unresectable HCC. BCLC = Barcelona Clinic Liver Cancer, AER = arterial enhancement ratio, TACE = transarterial chemoembolization, HCC = hepatocellular carcinoma Discussion This study represents the first application of habitat imaging technology to predict the occurrence of TACE refractoriness in patients with unresectable HCC. The entropy value of habitat 2, obtained through mode 2, was confirmed as a risk factor for TACE refractoriness. Based on this parameter, we integrated BCLC stage and MRI finding of blood products in mass to construct a predictive model, which was visualized through a nomogram. The model demonstrated favorable performance in both the training and validation cohorts. To summarize, our findings indicate that habitat analysis incorporating the AER map may serve as a promising tool for preoperatively identifying patients who are more prone to TACE failure, potentially guiding individualized treatment decisions and avoiding liver function impairment caused by ineffective embolization. Our study successfully employed the AER to characterize intratumoral heterogeneity in HCC. By comparing two habitat encoding strategies, we found that only when AER map was incorporated for tumor segmentation did the derived habitat feature demonstrate predictive value for TACE refractoriness, whereas mode 1 (based on T1-pre imaging) failed to establish correlation. This may be interpreted as when assessing HCC heterogeneity, T1-pre imaging mainly reflects the macroscopic composition of tissues, such as water distribution or fibrous components, and thus struggles to capture perfusion characteristics related to TACE efficacy. In contrast, the AER may address this limitation due to the following mechanisms: First, the development and progression of HCC are highly dependent on abnormal arterial blood supply [30], which is characterized by a large number of immature arteries with structural disorders and dysfunction. Such blood vessels are intuitively manifested as APHE on imaging [31]. Secondly, this kind of enhancement is often heterogeneous in HCC, and existing studies have confirmed that the “non-uniform” enhancement pattern of “APHE proportion < 50%” is closely associated with aggressive biological behaviors and poor prognosis of the disease [23, 32]. Furthermore, as a quantitative extension of APHE, the AER not only objectively quantifies the overall blood supply of the tumor [18], but more importantly, displays enhancement degree differences across various intratumoral regions at the voxel level, thereby enabling spatial visualization of perfusion heterogeneity. Consequently, the spatial distribution heterogeneity presented by AER may essentially depict the distributional disparities of abnormal neovessels within the tumor, establishing it as a key imaging biomarker for characterizing perfusion heterogeneity in HCC. Our study confirms that the AER entropy of habitat 2 is a key predictor of TACE refractoriness. This conclusion is supported by two main findings: on the one hand, compared to the average perfusion level provided by SI, the perfusion heterogeneity quantified by entropy better reflects tumor heterogeneity; on the other hand, the entropy of habitat 2 may possess unique biological significance. Specifically, although habitat 1 exhibits the highest AER entropy, it is habitat 2 with moderate AER entropy that shows a significant association with TACE refractoriness. This indicates that the spatial origin of heterogeneity holds greater clinical value than its numerical magnitude. We speculate that the high entropy in habitat 1 may primarily stem from signal disruption caused by necrosis or hemorrhage. In contrast, habitat 2 demonstrates a combination of “moderate entropy with low SI,” potentially characterizing a high-risk habitat of “low perfusion but metabolically active.” This area may exhibit active neovascularization that is poorly embolisable, making it prone to residual tumor cells and ultimately leading to local recurrence after TACE [33]. Previous studies have also shown that the aggressiveness of HCC is not always correlated with high perfusion, which is consistent with our findings [34, 35]. This suggests that certain specific low-perfusion areas may be critical drivers of treatment failure or disease progression. Compared with previous studies [7–10, 28, 36, 37] that have reported multiple risk factors for TACE refractoriness, such as tumor number, size, hypovascular nodules, BCLC stage, and AFP, our results showed that only BCLC stage maintained independent predictive value, while other clinical or imaging factors did not reach significance. This discrepancy may be attributed to the sample capacity limitations of our study. It is worth emphasizing that our study is the first to identify “blood products in mass” as a risk factor for TACE refractoriness. The underlying mechanism of this association might be closely related to the unique microenvironmental characteristics of hemorrhagic areas. These regions are often characterized by specialized conditions such as hypoxia and acidosis, which not only facilitate the enrichment and survival of cancer stem-like cells but may also induce the release of immunosuppressive factors, thereby weakening the anti-tumor immune response after TACE and collectively leading to treatment failure [38]. The prediction model based on AER demonstrated favorable predictive performance in both the training and validation cohorts. To enhance its clinical utility, we developed a nomogram for visual representation of the model. Clinicians may estimate the probability of TACE refractoriness for individual patients based on their specific BCLC stage and imaging characteristics. This visualization approach could offer an objective and quantitative basis for decision-making, possibly helping to develop individualized treatment strategies, such as prioritizing TACE or selecting alternative therapies, thereby improving treatment benefits. Our study has some limitations. First, while this study was specifically designed to characterize perfusion heterogeneity using the AER, future research may build upon this foundation by incorporating radiomics or deep learning-derived features into the habitat definition pipeline, either as additional inputs for clustering or as high-dimensional outputs from the subregions, to determine whether capturing biological information beyond perfusion can further enhance predictive performance. Second, as a single-center retrospective study, external validation in multi-center or prospective cohorts is needed to enhance reliability and applicability of the result. Finally, The biological mechanisms underlying the association between AER entropy and TACE refractoriness still require further elucidation through integrated molecular pathology or spatial transcriptomics. In summary, through dual-mode habitat encoding, our study is the first to confirm that habitat analysis using the AER to characterize intratumoral heterogeneity has potential utility for the preoperative non-invasive assessment of TACE refractoriness risk in patients with unresectable HCC. This approach holds promise for assisting clinicians in selecting optimal treatment strategies. Electronic Supplementary Material Below is the link to the electronic supplementary material. Supplementary Material 1 Acknowledgements Not applicable. Author contributions Conception and design: All authors. Final approval of the version to be submitted: All authors. Data acquisition: YX, WLS, QY. Data analysis: YX, WLS, HL, QY. Methodology: YX, WLS, HL, JZ, DJG. Statistical analysis: YX, HL. Resources: DJG. Supervision: DJG. Writing - original draft: YX. Writing - review and editing: DJG, WLS. Funding This research did not receive any specific grant from funding agencies in the public, commercial, or not-for-profit sectors. Data availability The datasets used and/or analyzed during the present study are available from the corresponding author on reasonable request. Declarations Ethics approval and consent to participate This study was approved by the Ethics Committee of the Second Affiliated Hospital of Chongqing Medical University, which waived the requirement for written informed consent owing to the use of deidentified retrospective data. The study was conducted in accordance with the Declaration of Helsinki. Consent for publication Not applicable. Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Yan Xiao and Wenlong Song contributed equally to this work. References Associated Data Supplementary Materials Supplementary Material 1 Data Availability Statement The datasets used and/or analyzed during the present study are available from the corresponding author on reasonable request.