Multi-scale and multi-parametric radiomics of gadoxetate disodium–enhanced MRI predicts microvascular invasion and outcome in patients with solitary hepatocellular carcinoma ≤ 5 cm

Objectives To develop radiomics-based nomograms for preoperative microvascular invasion (MVI) and recurrence-free survival (RFS) prediction in patients with solitary hepatocellular carcinoma (HCC) ≤ 5 cm. Methods Between March 2012 and September 2019, 356 patients with pathologically confirmed solitary HCC ≤ 5 cm who underwent preoperative gadoxetate disodium–enhanced MRI were retrospectively enrolled. MVI was graded as M0, M1, or M2 according to the number and distribution of invaded vessels. Radiomics features were extracted from DWI, arterial, portal venous, and hepatobiliary phase images in regions of the entire tumor, peritumoral area ≤ 10 mm, and randomly selected liver tissue. Multivariate analysis identified the independent predictors for MVI and RFS, with nomogram visualized the ultimately predictive models. Results Elevated alpha-fetoprotein, total bilirubin and radiomics values, peritumoral enhancement, and incomplete or absent capsule enhancement were independent risk factors for MVI. The AUCs of MVI nomogram reached 0.920 (95% CI: 0.861–0.979) using random forest and 0.879 (95% CI: 0.820–0.938) using logistic regression analysis in validation cohort (n = 106). With the 5-year RFS rate of 68.4%, the median RFS of MVI-positive (M2 and M1) and MVI-negative (M0) patients were 30.5 (11.9 and 40.9) and > 96.9 months (p < 0.001), respectively. Age, histologic MVI, alkaline phosphatase, and alanine aminotransferase independently predicted recurrence, yielding AUC of 0.654 (95% CI: 0.538–0.769, n = 99) in RFS validation cohort. Instead of histologic MVI, the preoperatively predicted MVI by MVI nomogram using random forest achieved comparable accuracy in MVI stratification and RFS prediction. Conclusions Preoperative radiomics-based nomogram using random forest is a potential biomarker of MVI and RFS prediction for solitary HCC ≤ 5 cm. Key Points • The radiomics score was the predominant independent predictor of MVI which was the primary independent risk factor for postoperative recurrence. • The radiomics-based nomogram using either random forest or logistic regression analysis has obtained the best preoperative prediction of MVI in HCC patients so far. • As an excellent substitute for the invasive histologic MVI, the preoperatively predicted MVI by MVI nomogram using random forest (MVI-RF) achieved comparable accuracy in MVI stratification and outcome, reinforcing the radiologic understanding of HCC angioinvasion and progression. Supplementary Information The online version contains supplementary material available at 10.1007/s00330-020-07601-2.

Microvascular invasion (MVI), present in 15-57.1% surgical specimens of HCC [7], is a well-established risk factor for postoperative recurrence [8,9], even for solitary small HCC [10]. To improve the prognosis of MVI-positive patients, a wide resection margin is recommended [11]. Therefore, preoperative diagnosis of MVI is of great importance for treatment strategies.
MVI is defined as the cancer cell nest in vessels lined with endothelium, which is visible only on microscopy [7,12] and poses a challenge for non-invasive diagnosis. Recently, preoperatively radiologic hallmarks including non-smooth tumor margin, peritumoral enhancement on arterial phase (AP), and peritumoral hypointensity on hepatobiliary phase (HBP) have shown to be conducive to MVI diagnosis but be inferior to radiomics signatures [13]. As a novel and non-invasive tool, radiomics can high-throughput extract quantitative imaging signatures to improve diagnostic or prognostic accuracy [14], which is also applicable to preoperative MVI and outcome prediction. Being related with postoperative recurrence and metastasis, peritumoral area of HCC is rich in highly invasive cells and susceptible to the formation of MVI [12], where it has been neglected in previous radiomics studies [11,15,16]. While gadoxetate disodium-enhanced (Gd-EOB-DTPA) MRI offers the identifiability of small or early HCC and the information of tumor heterogeneity and vascularization [17], previous radiomics studies [11,13] mainly focused on HBP images for predicting MVI. Thus, it is reasonable to investigate whether radiomics signatures extracted from intratumoral and peritumoral regions on multi-parametric images of Gd-EOB-DTPA MRI may allow more effective MVI prediction.
This study aimed to develop and validate nomograms based on multi-scale and multi-parametric radiomics of Gd-EOB-DTPA MRI for the preoperative MVI and outcome prediction in patients with solitary HCC ≤ 5 cm.

Study design and patients
Our hospital ethics committee approved this retrospective study and waived patient informed consent. Between March 2012 and September 2019, 356 pathologically confirmed HCC patients (303 males and 53 females; 54.22 ± 11.40 years) with preoperative Gd-EOB-DTPA MRI met the inclusion criteria ( Fig. 1): (a) solitary HCC with the longest diameter ≤ 5 cm; (b) without gross vascular invasion, bile duct tumor thrombosis or extrahepatic metastasis upon preoperative imaging; (c) without previous history of HCC-related treatments (hepatectomy, liver transplantation, chemotherapy, radiotherapy, transarterial chemoembolization, radiofrequency ablation, and immunosuppressive therapy); (d) complete histopathologic description of HCC; (e) MRI with sufficient image quality scanned within 1 month before surgery.
HCC pathological samples were taken by a 7-point baseline sample collection protocol [12]. Histopathological characteristics (tumor size, number, Edmondson-Steiner grade, MVI status and category, liver fibrosis grade based on the Scheuer scoring system, and Ki-67 protein expression) were assessed in consensus by two experienced abdominal pathologists.
MVI was defined as the presence of tumor in the portal vein, hepatic vein, or a large capsular vessel of the surrounding hepatic tissue lined with endothelium, which was visible only on microscopy [7,11,12,18]. According to the high-risk factors of adverse outcomes [12,18,19], the patients were classified into M0 (no MVI), M1 (invaded vessels were no more than five and located at the peritumoral region adjacent to the tumor surface within 1 cm), or M2 (MVI of > 5 or at > 1 cm away from the tumor surface) grades [12], respectively.

Radiomics analysis
Radiomics was implemented by Python programming language (version 3.7.3, https://www.python.org) with Pyradiomics (version 2.2.0, https://pyradiomics.readthedocs. io/en/latest/index.html) and Scikit-learn (version 2.1.0, https:// scikit-learn.org/stable/index.html) packages. Radiomics workflow comprised manual tumor segmentation, feature extraction and selection, multiple sequences and volumetric interests (VOIs) fusion, and model construction and evaluation (Fig. 2). First, tumor boundaries were manually delineated on all single sequence images, denoted as VOI tumor , by two radiologists (H.H.C. and L.Y., 8 years and 10 years of abdominal imaging experiences) with ITK-SNAP software (http://www. itksnap.org/pmwiki/pmwiki.php). Besides, the two radiologists randomly picked 5 to 10 blocks distributed in different liver lobes sufficiently away from large vessels, artifacts, liver margins, and hepatic lesions, which were used as regions of normal liver tissues (VOI liver ) for contrast analysis with tumor. To further explore the intratumoral and peritumoral information, the VOI tumor was shrunk 50% (VOI 50% ) and dilated by 5 mm and 10 mm (VOI 5mm and VOI 10mm ) using standard image morphological erosion and dilation operations, respectively. Please note that VOI 5mm and VOI 10mm excluded the tumor region and only referred to the peritumoral zone within 5 mm and 10 mm from the tumor surface. Meanwhile, a variety of regional combinations were experimented, including VOI tumor + liver which combined the tumor (VOI tumor ) and the liver background (VOI liver ) regions; VOI tumor + 5mm and VOI tumor + 10mm joined the tumor (VOI tumor ) with peripheral zones VOI 5mm and VOI 10mm , respectively, based on which VOI tumor + 5mm + liver or VOI tumor + 10mm + liver was defined with additional VOI liver merged.
Subsequently, a set of 854 features radiomics features were extracted from the original and three-dimensional wavelet filters images [24], including tumor shape, size, intensity, and texture (Table S2). These features were first selected by the Least Absolute Shrinkage and Selection Operator (LASSO ,  Table S3) for each VOI of each single sequence. The first selected features were then combined to obtain the optimal multi-VOI models in single sequences (Tables 2 and S4). These multi-VOI features of each sequence were finally joined and selected using LASSO again (Table S5) to derive the ultimate multi-sequence, multi-VOI radiomics model (Table S6), and based on which MVI nomograms were constructed with random forest (RF) and logistic regression (LR) classifiers respectively for comparison. Finally, the receiver operating characteristic, calibration, and decision curves were plotted and the validation data was tested for model evaluation.

Outcome analysis
Follow-up was performed at intervals of 3 to 6 months after curative surgery. The date of surgery, recurrence, metastasis, death, and the last follow-up were recorded for calculating the overall and recurrence-free survival (RFS).

Statistical analyses
Statistical analyses were performed with IBM SPSS Statistics (version 25) and R (version 3.6.1, https:// www.r-project.org) software. Patients enrolled in MVI or outcome study were randomly allocated to training a n d v a l i d a t i on c o h o r t s i n a ra t i o o f 7 : 3 . T h e discrimination performance of models was quantified by area under the curve (AUC) and net reclassification index (NRI). NRI > 0 meant a positive improvement, indicating that the predictive ability of the new model precedes the old one. Compared to the histologic MVI, the preoperatively predicted MVI status was calculated by MVI nomogram using RF (MVI-RF) or LR (MVI-LR) in each patient, with prediction probabilities > 50% classified into MVI-positive group and > 90% defined as M2 grade. P < 0.05 was considered statistically significant.
More details (T1 maps and morphologic hallmarks, feature extraction and selection, and statistical analyses) are available in the Supplementary Materials and Methods.     Fig. 3. The MVI predictive performances of clinical and imaging models are summarized in Table 3.

Performance of radiomics features from single sequences
The AUCs of each VOI in single sequences are displayed in Table 2. For the vast majority of VOIs and sequences, RF outperformed LR classifier, HBP, and PVP were superior to other sequences, and the VOI tumor + 10mm + liver yielded the best multi-VOI fusion for predicting MVI. Interestingly, the AUCs of VOI 50% , VOI tumor , VOI tumor + 5mm , VOI tumor + 10mm, and VOI tumor + 10mm + liver subgroups approximately kept increasing almost in all sequences regardless of the choice of classifiers. Notably, VOI tumor + 10mm + liver showed consistent performance improvements compared to VOI tumor in HBP and PVP sequences on the validation cohort (NRIs > 0, Table 4).

Performance of radiomics features from multiple sequences
In the VOI tumor subgroup, the MVI predictive efficacies of two best single sequences (HBP and PVP) were worse than any of the multi-sequence models either using RF or LR (Table S7), especially inferior to that of the best combination (PVP, HBP, AP, and pre-contrast T1 maps; AUCs of validation cohort: 0.871 using RF and 0.792 using LR; Fig. 4). Concretely, this four-sequence model showed significant improvements compared to the two best single sequences (HBP: NRI 19.28%, p = 0.046; PVP: NRI 20.90%, p = 0.017; Table 4 ) in the validation cohort using RF.
In the VOI tumor + 10mm + liver subgroup, the optimal multisequence fusion was the integration of PVP, HBP, AP, and DWI, with AUCs of 0.918 using RF and 0.809 using LR in the validation cohort (Tables 3 and S6). Meanwhile, the predictive performance of this four-sequence fusion-the final radiomics model-was also significantly superior to those of the two best single sequences (HBP: NRI 19.44%, p = 0.008; PVP: NRI 24.54%, p = 0.003; Table 4) as well as those of the clinical (NRI 54.1%, p = 0.002) and imaging models (NRI 22.2%, p = 0.029) using RF in the validation cohort. The details of the top six most discriminating features in the final radiomics model are provided in Table S8.

Performance of MVI nomograms
Based on the clinical, imaging, and final radiomics predictors, the ultimate MVI predictive model incorporated the independent risk factors of TBIL > 20.4 μmol/L, AFP > 20 ng/mL, incomplete or absent capsule enhancement, peritumoral enhancement, and higher score of radiomics (R-score) into visualized nomograms (Fig. 5a-b) as follows:   The predictive performances of MVI nomograms (Tables 3  and 4) demonstrated moderately or dramatically enhancements compared to those of clinical models (NRIs: 14.6-56.8%), imaging models (NRIs: 9.1-78.9%), and radiomics model using LR (NRIs: 19.4-35.7%), with a slightly negative improvement contrasted to radiomics model using RF (NRI < 0, p > 0.05). Furthermore, the AUCs (Fig. 4) of HBP sequence in VOI tumor , PVP sequence in VOI tumor , multisequence fusion in VOI tumor , the final radiomics in VOI tumor + 10mm + liver , and the ultimate predictive model of MVI (MVI nomogram) presented a gradual upward trend in validation cohort using RF or LR classifier. Being highly consistent with Fig. 3 Representative images of MVI-positive and MVI-negative patients. MVI-positive case: A 51-year-old male with elevated AFP, TBIL, and AKP levels (320 ng/mL, 32.6 μmol/L, and 131 U/L) was admitted to our department for abdominal discomfort and yellow sclera and identified intrahepatic recurrence at 11 months after hepatectomy. Gd-EOB-DTPA MRI detected a solid lesion (2.9 × 1.9 cm) in hepatic segment V, with the architectures of wedge-shaped peritumoral enhancement on arterial phase images (a, arrows), absent capsule enhancement on transitional phase images (b, arrows), non-smooth tumor edge on HBP, DWI, and HBP T1 maps (c-e, arrows), and typical MRI pattern of HCC (non-rim arterial phase enhancement and non-peripheral transitional phase hypointensity). M2 grade was diagnosis by postoperative pathological specimens with standard hematoxylin and eosin (HE, × 100): multiple tumor thrombi of microvasculature (f, black arrow; MVI > 5) were distributed in the widespread inflammatory cells, which were located at the region between the normal liver tissue in the right side and the infiltrating HCC lesion without tumor capsule in the upper left corner. MVI-negative case: A 77-year-old male with normal levels of AFP, TBIL, and AKP (3.4 ng/mL, 11.7 μmol/L, and 90 U/L) was admitted to our hospital for a liver lesion in health examination, and identified recurrence-free until April 2020 (18 months after hepatectomy). Gd-EOB-DTPA MRI detected a well-circumscribed solid lesion (2.3 × 2.0 cm) in hepatic segment II, with the architectures of absent peritumoral enhancement (g, arrows), intact capsule enhancement (h, arrows), smooth tumor margin (i-k, arrow), and typical MRI pattern of HCC. M0 grade was diagnosed by pathologic HE (× 100) sample: no tumor thrombus was detected in microvascular system (l, black arrow), which were located at the region between the normal liver tissue in the lower left corner and the HCC lesion with intact capsule in the upper right corner the actual MVI status in the calibration curves ( Fig. S1a-d), MVI nomograms obtained the best net clinical benefit, followed by the radiomics and imaging models, with the clinical model worst in the decision curves (Fig. S1e-h).
The results of multivariate Cox regression (Table 5)

Discussion
Our study developed radiomics-based nomograms for preoperatively predicting MVI and outcome in patients with solitary HCC ≤ 5 cm. The results demonstrated that AFP > 20 ng/mL, TBIL > 20.4 μmol/L, absent or incomplete capsule enhancement, peritumoral enhancement, and higher Rscore were independent risk factors for MVI. Mainly based on radiomics signatures of PVP, HBP, AP, and DWI in VOI tumor + 10mm + liver , the nomogram using RF or LR excellently identified MVI-positive patients. Furthermore, histologic MVI, ALT > 50 U/L, AKP > 125 U/L, and the elderly independently impaired RFS, with a relatively favorable prediction for recurrence. Histologic M0, M1, and M2 grades were significantly inverse correlated with RFS. Intriguingly, contrasted to histologic MVI, MVI-RF achieved comparable accuracy in MVI stratification and prognostic analyses. Elevated AFP level [7,11,16], incomplete capsule enhancement [7,8], and peritumoral enhancement [8,9,11] have been reported to be independent risk factors for MVI, which are consistent with our results. Independently facilitating MVI in this study, elevated TBIL level may secondary to the existence or obstruction of MVI in the biliary system [12,18,25]. This is partly due to the fact that cancerous thrombus in the newly formed bile ducts of tumor capsule [26], bile canaliculus, or interlobular bile ducts, rather than in gross or intrahepatic bile ducts, are difficult to be identified by preoperative imaging and excluded from the study population.
Peritumoral tissue is the first and most frequently vulnerable to MVI [21,27], the vessels of which further serve as the main hematogenous dissemination pathway of portal vein tumor thrombosis and metastasis [21]. Therefore, we constructed multi-VOI models for exploring this highly aggressive region. Interestingly, the AUCs of VOI 50% , VOI tumor , VOI tumor + 5mm , VOI tumor + 10mm , and VOI tumor + 10mm + liver signatures approximately kept increasing almost in all sequences irrespective of classifiers. These preponderances of VOI tumor over VOI 50% and VOI tumor + 10mm over VOI tumor for predicting MVI were consistent with the CT results of Xu et al [8] and HBP results of Feng et al [13], respectively. Meanwhile, the AUCs of VOI tumor , VOI tumor + Fig. 4 Receiver operating characteristic curves of different models for predicting MVI. Receiver operating characteristic curves of different models for predicting MVI were plotted by random forest (a: training cohort, b: validation cohort) and logistic regression (c: training cohort, d: validation cohort) to crossly validate the robustness of models liver and VOI tumor + 5mm + liver (VOI tumor + 10mm + liver ) features, as well as those of VOI tumor , VOI tumor + 5mm (VOI tumor + 10mm ) and VOI tumor + 5mm + liver (VOI tumor + 10mm + liver ) signatures, also showed an increasing trend. Notably, the performance of VOI tumor + 10mm + liver signatures preceded that of VOI tumor + 10mm features either in this paper or in Feng et al study [13]. Besides, the optimal multisequence fusion outperformed the two best single sequences both in VOI tumor and in VOI tumor + 10mm + liver subgroups. These results signified the superiority of tumor periphery compared with tumor interior, the significance of texture and intensity difference between normal liver and intra-/peritumoral tissue, and the synergistic effect of multisequence and multi-VOI fusion for predicting MVI, which have been neglected in and might be the reason why our MVI nomograms obtained better performances than previous radiomics studies [8,11,13,15,16,28].
Likewise, the top 6 most discriminating signatures of the final radiomics model also indicated the importance of peritumoral and intratumoral fusion. Being partly coincided with previous studies [8,11], the six signatures included tumor size, shape, and intratumoral and peritumoral texture heterogen e i t y . B y d e f i n i t i o n , H B P _ V O I 5 m m _ w a v e l e t -HHL_firstorder_Energy and HBP_VOI 5mm _wavelet-HLL_glszm_GrayLevelNonUniformity involved the texture heterogeneity of the peritumoral tissue within 5 mm, which might reflect an aggressive tendency to invade the tumor capsule and protrude into the peritumoral non-neoplastic parenchyma [27]. In addition, HBP_VOI tumor _original_shape_Sphericity and DWI_VOI tumor _original_shape_MajorAxisLength represented the spherical disproportion and the largest axis length of tumor, respectively. These were analogue to the well-known independent hallmarks "non-smooth edge and the longest diameter of tumor" of MVI [7][8][9]11]. Furthermore, HBP_VOI tumor _wavelet-H L L _ g l s z m _ S i z e Z o n e N o n U n i f o r m i t y a n d HBP_VOI tumor _original_glszm_GrayLevelNonUniformity concerning intratumoral texture heterogeneity might be induced by tumor cellularity, micronecrosis and inflammation, for which further facilitated MVI [11,29]. Coincidentally, five-sixths features were extracted from HBP, suggesting the significance of Gd-EOB-DTPA MRI in MVI diagnosis. Histologic MVI [8,10,30], the elderly [30][31][32], incomplete or absent capsule enhancement [33,34], and elevated ALT [30][31][32] and AKP [35][36][37] levels have been reported to impair outcomes of HCC patients, which were corresponded to our results. Conforming to the outcomes of few studies with histologic MVI grades [19,38], our histologic MVI stratification, especially the novel and non-invasive MVI-RF classifications, showed significantly inverse correlations with RFS. Hence, the MVI-RF-an excellent substitute of histologic MVImay be employed in patients with solitary HCC ≤ 5 cm, especially for those who underwent ablation without histologic MVI data. Namely, MVI-positive or even M2-grade patients diagnosed by MVI-RF before ablation might require more active clinical treatment and intense follow-up. Nevertheless, the AUCs of RFS nomograms around 0.66 for histologic MVI and MVI-RF subgroups, the unsatisfactory results may be induced by (1) the paucity of postoperative characteristics (e.g., preventive transarterial chemoembolization, immunosuppression therapy); (2) the absence of robust radiomics analysis in terms of recurrence instead of MVI; (3) the exclusion of well-established key predictors of recurrence (e.g., tumor size beyond 5 cm, satellite nodules or multifocal HCC, cancerous thrombus in gross bile ducts or vessels) in our study population.
This study has several limitations. Firstly, this paper is a retrospective single-centre study in China and needs to be validated by the external cohort. Secondly, we did not incorporate genomics with radiologic hallmarks, just as Banerjee et al [39]. Thirdly, this study focused on the solitary HCC within 5 cm, leading to a slightly lower frequency of MVI in our population than those of previous MVI studies with macrovascular invasive, larger, or multifocal HCC [7][8][9]. Fourthly, the radiomics results may slightly vary between different radiomics or statistical analysis software from feature selection to model evaluation. Hence, the well-recognized LASSO algorithm of R software [13,40], Pyradiomics [40][41][42], and Scikitlearn [43,44] packages of Python software were also employed to this paper, for facilitating the future study to verify the robustness of our findings. Finally, HCC Fig. 6 Kaplan-Meier curves of recurrence-free survival. With the Kaplan-Meier analysis and 2-sided log-rank test, recurrence-free survival curves were scaled by the histologic MVI status (a) and the predicted MVI status (b) by MVI nomogram using random forest (MVI-RF) and were further stratified by the histologic MVI (c) and MVI-RF grades (d), respectively has a strong male preponderance [45], and thus, the sex ratio imbalance-the inherent selection bias-cannot be completely avoided in this study.
In summary, mainly based upon multi-parametric radiomics in VOI tumor + 10mm + liver of Gd-EOB-DTPA MRI, the nomogram using random forest is a potential biomarker for preoperatively predicting MVI and RFS in patients with solitary HCC ≤ 5 cm.