Predictive value of a constructed artificial neural network model for microvascular invasion in hepatocellular carcinoma: A retrospective study

Abstract

BACKGROUND

Microvascular invasion (MVI) is a significant risk factor for recurrence and metastasis following hepatocellular carcinoma (HCC) surgery. Currently, there is a paucity of preoperative evaluation approaches for MVI.

AIM

To investigate the predictive value of texture features and radiological signs based on multiparametric magnetic resonance imaging in the non-invasive preoperative prediction of MVI in HCC.

METHODS

Clinical data from 97 HCC patients were retrospectively collected from January 2019 to July 2022 at our hospital. Patients were classified into two groups: MVI-positive (n = 57) and MVI-negative (n = 40), based on postoperative pathological results. The correlation between relevant radiological signs and MVI status was analyzed. MaZda4.6 software and the mutual information method were employed to identify the top 10 dominant texture features, which were combined with radiological signs to construct artificial neural network (ANN) models for MVI prediction. The predictive performance of the ANN models was evaluated using area under the curve, sensitivity, and specificity. ANN models with relatively high predictive performance were screened using the DeLong test, and the regression model of multilayer feedforward ANN with backpropagation and error backpropagation learning method was used to evaluate the models’ stability.

RESULTS

The absence of a pseudocapsule, an incomplete pseudocapsule, and the presence of tumor blood vessels were identified as independent predictors of HCC MVI. The ANN model constructed using the dominant features of the combined group (pseudocapsule status + tumor blood vessels + arterial phase + venous phase) demonstrated the best predictive performance for MVI status and was found to be automated, highly operable, and very stable.

CONCLUSION

The ANN model constructed using the dominant features of the combined group can be recommended as a non-invasive method for preoperative prediction of HCC MVI status.

Key Words: Hepatocellular carcinoma; Texture analysis; Magnetic resonance imaging; Microvascular invasion; Pseudocapsule; Tumor blood vessels

Core Tip: Whether artificial neural network (ANN) models mimicking human brain provide a high predictive value for hepatocellular carcinoma (HCC) microvascular invasion (MVI). The preoperative prediction ANN model constructed on multiparametric magnetic resonance imaging is acute and stable. ANN models have clinical values and benefits in non-invasively predicting MVI in HCC patients.

INTRODUCTION

Primary liver cancer is one of the most common malignant tumors worldwide, with hepatocellular carcinoma (HCC) comprising approximately 85%-90% of all primary liver cancer cases[1]. Although surgical resection remains a primary treatment modality for HCC, the postoperative recurrence rate continues to be notably high[2]. Microvascular invasion (MVI) is recognized as a significant risk factor for tumor recurrence and metastasis following HCC surgery[3-5]. Currently, the evaluation criteria for MVI predominantly rely on postoperative pathological biopsy, and there is a notable lack of preoperative evaluation methods. Imaging texture analysis has been applied to various medical imaging settings to assist in disease diagnosis and treatment[6-9], as it can provide objective and quantitative image description features that reflect the physiological heterogeneity within the region of interest (ROI). Previous studies have often utilized computerized tomography (CT) images[10-13] and two-dimensional ROIs, which provide less information compared to three-dimensional (3D) ROIs[14-16]. Multiparametric liver magnetic resonance imaging (MRI) offers advantages in assessing tumor heterogeneity and has been employed for clinical diagnosis, staging classification, and efficacy evaluation of liver cancer[17,18]. However, previous investigations into the predictive performance of MRI imaging features for MVI have yielded inconsistent results and are highly dependent on the diagnosing physicians. Artificial neural networks (ANNs) are mathematical models that emulate the functions of the human brain. Consisting of a large number of interconnected neurons, ANNs exhibit high fault tolerance, parallel-distributed processing ability, adaptability, self-organization, and self-learning abilities, which enabling them to analyze relatively complex nonlinear systems[19]. This study aims to address the question: Will an ANN model that incorporates texture features from multiparametric MRI images, in conjunction with radiological signs, demonstrate a higher predictive value for HCC MVI? We hypothesize that the integration of artificial intelligence and traditional imaging features could represent a significant advancement in predicting MVI, thereby enhancing clinical applications.

MATERIALS AND METHODS

Study subjects

In accordance with the ethical principles of the Declaration of Helsinki, we conducted a retrospective analysis involving 97 patients with HCC, all of whom were confirmed by surgical pathology at our hospital between January 2019 and July 2022. This study received approval from our institutional review board, and informed consent was obtained from all patients (No. 2015-02-28-1).

Research methods

MVI assessment and grouping: MVI refers to the presence of cancer cell clusters within vascular lumens lined by endothelial cells, as detected under the microscope. Liver cancer is most commonly found to invade the branches of the portal vein (including intra-capsular blood vessels)[20]. Based on postoperative pathological observations, patients were classified into MVI-positive and MVI-negative groups.

MRI examination method: (1) MRI examination equipment: UMR790 3.0T MRI and SIEMENS Verio 3.0T MRI imaging devices; (2) Contrast agents: Gadobutrol Injection (0.1 mmol/kg, injection rate 1.5 mL/second) or gadopentetate dimeglumine (0.2 mmol/kg, injection rate 2.0 mL/second); and (3) Scanning range and parameter settings: The scanning range was from the top of the diaphragm to the lower edge of the liver.

Scanning parameter settings were as follows: (1) UMR790 3.0T MRI: Axial t1 gre-quick3d-tra [time of repetition (TR)/time of echo (TE) 2.65/1.03 ms], slice thickness 3 mm; Axial t2-arms-tra-fs-navi (TR/TE 4352/92.4 ms), slice thickness 6 mm. The contrast agent was administered via the cubital vein using a dual-barrel high-pressure injector. Following the contrast agent injection, 20 mL of 0.9% sterile saline was immediately injected at the same rate to flush the tube. An automatic monitoring scanning method was then employed to capture six arterial phases (AP) at 15 to 28 seconds post-contrast agent injection, with selection of the optimal late AP image. Portal venous phase (VP) images were subsequently obtained at 55 to 65 seconds, and delayed phase (DP) images were captured at 180 seconds; and (2) SIEMENS Verio 3.0T MRI: Axial-t1-vibe-fs-tra (TR/TE 3.92/1.39 ms), slice thickness 4 mm; Axial t2-blade-tra-fs-navi (TR/TE 4185.31/89.00 ms), slice thickness 6 mm. The contrast agent was similarly injected via the cubital vein using a dual-barrel high-pressure injector, followed by the immediate injection of 20 mL of 0.9% sterile saline at the same rate to flush the tube. Late AP, portal VP, and DP images were obtained at 25 to 30 seconds, 55 to 65 seconds, and 180 seconds after contrast agent injection, respectively.

HCC MRI radiological sign evaluation

Two radiologists independently evaluated the radiological signs without knowledge of the pathological diagnosis of patients. In cases of inconsistent evaluation results, a consensus was achieved through discussion or by consulting another senior diagnostic physician. The imaging criteria included in this study are based on the standardized diagnosis and treatment quality control indicators for Chinese primary HCC (2022 Edition)[21]. Examples of several MRI radiological signs are illustrated in Figure 1.

Figure 1 Representative examples of hepatocellular carcinoma magnetic resonance imaging radiological signs. A: Axial portal venous phase image: A high-signal ring surrounding the tumor lesion edge in the left lobe of the liver, demonstrating good continuity, indicative of a complete pseudocapsule; B: Axial delayed phase image: A high-signal ring around the tumor lesion edge in the right lobe of the liver with interrupted continuity, suggesting an incomplete pseudocapsule; C: Axial T2-weighted imaging image: Irregular liver margin and diffuse reticular slightly low-signal change of the liver parenchyma, characteristic of cirrhosis; D: Axial arterial phase image: Thickened and tortuous enhanced blood vessels within the tumor lesion in the right lobe of the liver, representing tumor vessels; E: Axial T2-weighted imaging image: Irregular patchy high-signal region within a large tumor, suggesting areas of cystic degeneration or necrosis; F: Axial delayed phase image: Corresponding to E, showing no enhancement in the areas of suspected cystic degeneration or necrosis.

Grouping was based on several criteria, including maximum tumor diameter, pseudocapsule status, tumor blood vessels, cystic degeneration or necrosis, and cirrhosis: (1) Maximum tumor diameter: Lesions were measured using the medical imaging department’s picture archiving and communication system (PACS) and categorized into two groups: < 3.0 cm and ≥ 3.0 cm, according to the maximum diameter of the tumor; (2) Pseudocapsule status: Based on the T1-weighted imaging (T1WI) sequence showing a low signal ring of 0.5 mm-3 mm at the edge of lesion, or a high signal shadow with ring-shaped delayed enhancement during the portal VP or DP. The pseudocapsule status was classified as either complete or incomplete/absent; (3) Tumor blood vessels: Based on whether thickened and tortuous enhancing blood vessels within the tumor showed during the AP of the MR-enhanced scan, and categorized as either with or without tumor blood vessels; (4) Cystic degeneration or necrosis: The observation of patchy low signal intensity on T1WI and high signal intensity on T2-weighted imaging (T2WI) within the lesion, combined with the absence of enhancement on the contrast-enhanced scan, suggests the presence of necrosis or cystic degeneration. Conversely, in the absence of these findings, it was classified as no detected necrosis or cystic degeneration; and (5) Cirrhosis: The identification of an irregular liver edge, blunt liver margin, widened liver fissures, disproportionate liver lobes, and diffuse reticular low signal changes in the liver parenchyma on T2WI sequence indicated cirrhosis. In contrast, the absence of these features was interpreted as indicative of no cirrhosis.

Texture analysis method based on multiparametric MRI

HCC lesion ROI delineation and texture feature extraction: T2WI, AP, VP, and DP images were exported from the PACS in BMP format. Initially, the MaZda 4.6 software (available from: http://www.eletel.p.lodz.pl/mazda/) was employed to manually delineate the ROI layer by layer on the enhanced scanning phase images with the clearest tumor boundaries. These delineated layers were subsequently fused to create a 3D volume of interest (VOI). The VOI was then copied to other enhanced images, and in instances of misalignment, adjustments were made based on the neighboring tissue structure to minimize placement errors across different scanning phase images. Given that the layer thickness of T2WI images differed from that of the enhanced images, ROIs needed to be manually delineated separately, layer by layer, before being fused into a single VOI. Finally, a senior radiologist meticulously reviewed the placement of all VOIs layers, examining them layer by layer. In cases of disagreement regarding the ROI delineation or VOI adjustments, the lesion was re-delineated or adjusted until consensus was achieved. To minimize the impact of partial volume effects, ROIs were manually delineated on all slices with locating 1 mm-2 mm inside the tumor lesion boundary. Before extracting texture features, all enrolled MRI images were gray-level standardized to reduce the impact of imaging non-standardization on the stability of texture features[13,16]. A total of 94 texture feature parameters were extracted for each sequence separately.

Dimensionality reduction of HCC MRI texture features: The mutual information (MI) method included in the MaZda 4.6 software was used to select the top 10 dominant texture features from single sequence MRI (T2WI/AP/VP/DP) texture features. Python programmed MI was implemented for dimensionality reduction and selection of the top 10 dominant texture features from multiparametric MRI (T2WI + AP + VP + DP) texture features, totaling 94× 4 features.

Construction and evaluation of ANN models: Using the B11 program included in the MaZda 4.6 software, the dominant texture features obtained from dimensionality reduction of single sequence MRI and multiparametric MRI were used to construct ANN models separately. The ten dominant texture features from the multiparametric MRI were combined with radiological signs, which are independent predictors of HCC MVI, to develop the ANN model. The results were represented as misclassification rates (MCR). The predictive performance of the model for MVI was evaluated using the area under the receiver operating characteristic (ROC) curve (AUC), as well as sensitivity and specificity metrics. The DeLong test was employed to assess the statistical differences in AUC among the ANN models based on different features, leading to the selection of models that demonstrated relatively high performance. The flowchart illustrating the construction of the ANN models based on MRI texture features is presented in Figure 2.

Figure 2 Flowchart of artificial neural network model construction based on magnetic resonance imaging texture features. MRI: Magnetic resonance imaging; ANN: Artificial neural network; VOI: Volume of interest; MVI: Microvascular invasion.

Training and testing of ANN models with high prediction performance: A regression model based on multilayer feedforward backpropagation ANN and error backpropagation learning method was used for training and testing ANN models with high prediction performance. The ANN was implemented using the Python programming language, using features of ANN models demonstrating high prediction performance as input variables and postoperative pathological results of MVI-positive and MVI-negative groups as output variables. The dataset was divided into an 8:2 ratio using a random selection method, followed by input variable data standardization. The dense connection layer of the ANN model was configured with 256 neurons and utilized a sigmoid activation function, with input dimensions corresponding to the number of features. The hidden layer was composed of 128 neurons, also employing a sigmoid activation function. The optimizer used was root mean square propagation, and the loss function chosen was categorical cross entropy. The number of neurons in the hidden layer was determined based on the quantity of inputs in the input layer. During model training, the number of iterative calculations was set to 200, batch size was set to 64, and 30% of the data was used as a test set. The output layer consisted of two elements, with the first element representing patients in the MVI-positive group and the second element representing patients in the MVI-negative group. Finally, the accuracy rate and loss function values of the test set data pertaining to HCC MVI prediction were monitored, and the calculations were repeated 30 times. The average value obtained from these repetitions was considered the final result.

Statistical analysis

Statistical product and service solutions 26.0 software analysis package was used for this study, and two independent samples t-test was used for measures that conformed to normal distribution, and nonparametric Mann-Whitney U test was used for measures that did not conform to normal distribution; χ² test was used to analyze the relationship between MRI imaging signs and MVI expression; and for clinical or imaging features with P < 0.05, binary logistic regression was used to further screen for independent independent risk factors. Python 3.7 programming language with MI method was used to perform feature dimensionality reduction for multiparameter texture features and select the top 10 dominant features. The ANN model of the built-in B11 program of MaZda 4.6 software was used for prediction and classification. The performance of ANN models in predicting MVI status was represented by the MCR. The AUC, sensitivity, and specificity were metrics used to evaluate the predictive performance. To assess the statistical differences between the AUCs of different models, the DeLong test was applied, with a P value of less than 0.05 considered statistically significant.

RESULTS

Prediction of HCC MVI based on MRI radiological signs

Based on the postoperative pathological results, patients were divided into the MVI-positive group (57 cases in total, including 50 males and 7 females, aged 29-80 years, with an average age of 51.28 ± 12.54 years) and the MVI-negative group (40 cases in total, including 34 males and 6 females, aged 35-79 years, with an average age of 55.23 ± 11.33 years). The flowchart for the enrollment process of the study population is shown in Figure 3. The relationship between MRI imaging signs and HCC MVI positive and negative groups was analyzed using the χ² test, which are both categorical variables. Among the clinical and radiological signs including gender, age, tumor maximum diameter (≥ 3 cm or < 3 cm), pseudocapsule status (presence and complete or Absence/incomplete), tumor blood vessels (presence or absense), cystic degeneration or necrosis (presence or absense) and cirrhosis (presence or absense), it was found that tumor maximum diameter, pseudocapsule status, presence of tumor blood vessels, and presence of necrosis or cystic degeneration were significantly correlated with MVI status (all P < 0.05) (Table 1), while the presence of cirrhosis was not correlated with MVI (P = 0.298). For clinical or imaging features with significant correlation (P < 0.05), binary logistic regression was used to further screen independent risk factors. The results showed that the absence of pseudocapsule or incomplete pseudocapsule and tumor blood vessels were independent risk factors for MVI (Table 1).

Figure 3 Flowchart of subject enrollment in this study. MRI: Magnetic resonance imaging; MVI: Microvascular invasion; HCC: Hepatocellular carcinoma; ROI: Region of interest.

Table 1 Binary logistic regression analysis for microvascular invasion prediction.

Variables	β	SE	Wald	df	P value	OR (95%CI)
Tumor maximum diameter	-0.074	0.588	0.016	1	0.900	0.929 (0.293-2.942)
Pseudocapsule	2.093	0.639	10.733	1	0.001	8.111 (2.318-28.373)
Tumor blood vessels	2.051	0.788	6.775	1	0.009	7.775 (1.660-36.421)
Cystic degeneration or necrosis	0.098	0.677	0.021	1	0.885	1.103 (0.293-4.4155)

OR: Odds ratio; CI: Confidence interval.

Application and evaluation results of ANN models for MVI status prediction

The dominant texture features obtained from single sequence and multiparametric MRI after MI dimensionality reduction are shown in Table 2. The texture features of multiparametric MRI were derived from the gray-level gradients and long run length matrix features of AP and VP images. The performance of ANN models constructed on different features for MVI performance prediction is shown in Table 3, and the ROC curves are shown in Figure 4. Notably, the ANN model incorporating the combined group (pseudocapsule situation, tumor blood vessels, AP, and VP) exhibited the highest performance for predicting MVI, with metrics including MCR, sensitivity, specificity, and AUC recorded at 13.40%, 80.70%, 97.50%, and 0.891, respectively. Subsequently, a difference analysis of AUCs using ANN models constructed on dominant texture features from different MRI sequence and radiological signs were performed, and the results showed that the difference between the combined group and the single sequence or AP + VP ANN models was statistically significant (P < 0.05 for both comparisons). The ANN models constructed on T2WI, AP, VP, and AP + VP dominant texture features demonstrated comparable performance for predicting MVI, with no statistically significant difference between pairwise comparisons (P > 0.05 for all). These findings suggest that the ANN model utilizing features from the combined group demonstrates superior predictive performance, which warrants further testing and validation (the schematic diagram of the combined group ANN model structure is shown in Figure 5).

Figure 4 Receiver operating characteristic curves for predicting hepatocellular carcinoma microvascular invasion using artificial neural network models constructed on different sequence features. T2WI: T2-weighted imaging; AUC: Area under curve; AP: Arterial phases; VP: Venous phase; DP: Delayed phase.

Figure 5 Schematic diagram of the combined group artificial neural network model structure. AP: Arterial phases; VP: Venous phase; MVI: Microvascular invasion; HCC: Hepatocellular carcinoma.

Table 2 The dominant texture features selected after microvascular invasion dimensionality reduction.

T2WI	AP	VP	DP	Multiparametric
GrSkewness	Z_ShrtREmp	Z_LngREmph	Perc.01% 3D	AP-S (0, 1, 0) SumAverg
135dr_GLevNonU	Z_Fraction	Z_Fraction	S (0, 0, 1) InvDfMom	AP-Horzl_Fraction
45dgr_GLevNonU	Z_LngREmph	S (1, 0, 0) SumAverg	GrMean	AP-Horzl_ShrtREmp
Horzl_GLevNonU	45dgr_ShrtREmp	S (0, 0, 1) SumAverg	Z_LngREmph	AP-Horzl_LngREmph
Z_RLNonUni	S (0, 0, 1) InvDfMom	S (1, 1, 0) SumAverg	Z_ShrtREmp	AP-S (0, 0, 1) InvDfMom
135dr_ShrtREmp	S (0, 0, 1) SumAverg	Vertl_RLNonUni	S (1, -1, 0) SumAverg	VP- 45dgr_RLNonUni
Horzl_Fraction	Skewness 3D	135dr_RLNonUni	S (1, -1, 0) SumVarnc	VP-Vertl_RLNonUni
135dr_RLNonUni	GrSkewness	Z_ShrtREmp	Z_Fraction	VP-S (0, 0, 1) SumAverg
Z_GLevNonU	45dgr_Fraction	Perc.10% 3D	S (0, 1, 0) SumVarnc	VP-Z_ShrtREmp
Vertl_GLevNonU	Kurtosis 3D	S (0, 1, 0) SumAverg	Perc.10% 3D	VP-Z_LngREmph

T2WI: T2-weighted imaging; 3D: Three-dimensional; AP: Arterial phases; VP: Venous phase; DP: Delayed phase.

Table 3 Hepatocellular carcinoma microvascular invasion prediction results from artificial neural network models constructed on different features, n (%).

Sequence	MCR (n = 97)	Sensitivity (%)	Specificity (%)	AUC (95%CI)
T2WI	25 (25.77)	80.70	65.00	0.729
AP	19 (19.59)	100.00	52.50	0.762
VP	24 (24.74)	70.17	82.50	0.763
DP	23 (23.71)	70.17	85.00	0.776
AP + VP	17 (17.53)	94.73	65.00	0.799
1Combined	13 (13.40)	80.70	97.50	0.891

¹Combined = Pseudocapsule + tumor blood vessels + arterial phases + venous phase.

T2WI: T2-weighted imaging; MCR: Misclassification rate; AUC: Area under curve; AP: Arterial phases; VP: Venous phase; DP: Delayed phase; CI: Confidence interval.

Training and testing results for the ANN model of high prediction performance

In this study, the ANN model for the combined group was trained and tested, with the accuracy rate and loss function values for the test set being monitored. After 30 iterations, the mean value was obtained as the final result. The results show that the test set data had an accuracy rate of 0.774 ± 0.335 for predicting HCC MVI, and the loss function value was 0.647 ± 0.061. These results suggest that the model demonstrates good stability.

DISCUSSION

MRI radiological signs in predicting MVI

When the tumor diameter exceeds 3.0 cm, significant alterations occur in the genetic material and biological invasiveness of HCC[22]. Based on this theoretical foundation, this study used 3.0 cm as the threshold for classifying HCC size. Previous studies have found that HCC tumor size, multiple nodular tumor morphology, irregular tumor margins, incomplete pseudocapsule, and abnormal intratumoral arteries are radiological signs related to MVI, while their clinical values and applicability vary. Huang et al[23] found that tumor size and incomplete pseudocapsule are independent risk factors for predicting MVI. In contrast, this study found that while tumor size is correlated with MVI status, it was not identified as an independent risk factor for predicting MVI, aligning with the findings of Kim et al[24] and Chou et al[25], and different from Huang et al[23]. This discrepancy may be attributed to selection bias regarding HCC lesion size. Furthermore, this study identifies the absence or incompleteness of a pseudocapsule as an independent risk factor for predicting MVI, aligning with the findings of Huang et al[23]. This observation supports the hypothesis that the pseudocapsule of HCC serves to inhibit cancer cell invasion into adjacent liver parenchyma. Additionally, cancer cell cords at the incomplete pseudocapsule can interact with normal liver cells, with tumor blood vessels directly linking tonormal liver sinusoids, thereby increasing the likelihood of MVI by cancer cells[26]. In addition to these above findings, we further proved that tumor blood vessels are also an independent risk factor for predicting MVI, consistent with the finding of Renzulli et al[27]. Unlike Renzulli et al[27] who combined intratumoral arteries and low density halo sign into a specific radiological sign/imaging feature called two-trait predictor of venous invasion for preoperative prediction of MVI, this study did not include low signal halo analysis due to the difficulty in distinguishing low signal halo and enhanced pseudocapsule on MRI images. In the study of preoperative MVI risk assessment for HCC by Yang et al[28], least absolute shrinkage and selection operator (LASSO) regression analysis was performed on a cohort of 405 patients who underwent HCC resection. In this study, HCC tumor necrosis or cystic degeneration shows statistical significance in detecting difference between MVI positive and negative cases, however without being an independent risk factor for predicting MVI status. Cirrhosis is an important indicator of poor prognosis in HCC[29], and some studies have reported the presence of cirrhosis as an independent risk factor for MVI in HCC patients. In contrast to these previous finding, our study did not establish a correlation between the occurrence of MVI in HCC patients and a history of cirrhosis.

MVI prediction with ANN models constructed on MRI texture features

Tumors are heterogeneous entities that encompass diverse scales[30], and invasive biopsy-based detection methods have limitations and often fail to capture the overall biological characteristics of a tumor. Tumor texture features mainly include characteristics derived from gray-level histograms, gray-level co-occurrence matrices, gray-level gradients, and long-run matrices, which can reflect subtle texture changes that are challenging to detect with the naked eye. Some of these features may have significant implications for the tumor and its surrounding microenvironment[12,31]. In this study, similar to the research of Zhang et al[32] and Meng et al[33], 3D VOI image data analysis was used. Previous studies have established that variations in contrast agents, slice thickness, and machine acquisition parameters can affect the repeatability and variability of radiomics[12,34], leading to inconsistencies in pixel or voxel size, gray-level number and range. These factors may affect the performance of texture features, interfering with the texture analysis model’s applications in other centers. Therefore, it is essential to perform preprocessing with image resampling and/or gray-level normalization before texture features extractions. In this study, the images were gray-level standardized, minimizing the bias of texture features. Zhu et al[35] performed dimensionality reduction on the texture features of MRI AP and VP images and obtained four MR texture features of AP images and five texture features for portal venous phase (PP) images. They established individual logistic regression models utilizing texture features to predict MVI. In the ROC analysis, the AP texture model demonstrated superior diagnostic performance compared to the PP model in the validation cohort, with an AUC of 0.773 vs 0.62, which is similar to the AUC value of the ANN model constructed in this study based on AP image texture features. Zhu et al[35] focused merely on the texture features of arterial and portal VP, neglecting the predictive value of T2WI and DP image texture features, which have also been routinely used for preoperative MVI prediction. Previous studies[36,37] have shown that diffusion-weighted imaging (DWI)/T2WI mismatch and apparent diffusion coefficient (ADC) measurements are not particularly reliable for predicting MVI, so this study did not incorporate DWI/T2WI mismatch and ADC. Studies by Zhang et al[32], Guo et al[12], and Zhu et al[35] have demonstrated that texture analysis prediction models based on the MRI AP are superior to other single sequence models. Given that HCC increasingly receives blood supply from the hepatic artery, which ultimately predominates, the AP images are believed to more accurately reflect tumor heterogeneity and, consequently, better assess MVI. In this study, the ANN models constructed on single sequence MRI (T2WI/AP/VP/DP) and AP + VP dominant texture features had no statistically significant. This lack of significance may be attributed to variations in the dimensionality reduction methods applied to the texture features, as well as differences in the model construction approaches. Further studies with an increased sample size are necessary to confirm these findings and to determine whether the texture feature values obtained from routine MRI scans are comparable to those derived from enhanced image sources. It is worth mentioning that Nebbia et al[38] found that when tumor and margin radiomics features were combined, the performance of the model decreased, indicating that the information obtained from the margin and the tumor itself are not complementary (and might even be conflicting). Similarly, in this study, the prediction performance of the ANN model constructed on AP + VP dominant texture features for HCC MVI did not provide statistically significant additional value to the ANN model constructed on single sequence MRI dominant features. Xu et al[39] focused on the CT radiographic features around the tumor according to the MVI definition, but it was not superior to the features obtained from the tumor itself in terms of predicting MVI. Therefore, this study only used the entire tumor texture features to predict MVI and systematically evaluated the predictive performance of T2WI and three phase enhanced MRI images for MVI, all of which exhibited certain predictive capabilities for MVI status.

ANN model constructed on multiparametric MRI texture features combined with radiological signs for MVI prediction

Currently, there is an ongoing debate regarding the extent to which the combination of texture features or radiomics features with radiological signs can enhance the predictive performance of HCC MVI. In this study, the ANN model constructed using the dominant features from the combined group (pseudo-capsule status + tumor blood vessels + AP + VP) achieved a MCR, sensitivity, specificity, and AUC of 13.40%, 80.70%, 97.50%, and 0.891, respectively, for the high-expression group of HCC MVI. The accuracy rate of the test set data for HCC MVI prediction was 0.774 ± 0.335, with a loss function value of 0.647 ± 0.061, indicating good model stability. The differences between this model and the ANN models constructed on single sequence or AP+VP dominant features were statistically significant (P < 0.05 for both). In the ANN model, the pseudocapsule status and tumor blood vessels provided statistically significant additional value for the prediction of MVI based on texture features. Lu et al[40] studied the preoperative radiomics prediction of HCC MVI based on gadolinium ethoxybenzyl diethylenetriamine pentaacetic acid (Gd-EOB-DTPA)-enhanced MRI and established a combined model using tumor margin, peritumoral low intensity, and seven radiomics features. This combined model demonstrated superior performance compared to both the radiomics model and the clinical radiology model, achieving the highest sensitivity of 90.89% in the validation set. The AUC values for the combined model, radiomics model, and clinical radiology model were 0.826, 0.755, and 0.708, respectively. The AUC of the combined ANN model for predicting MVI in this study was slightly higher than that reported by Lu et al[40], Zhang et al[13], based on CT enhancement images of radiomics features in 637 patients, established a logistic regression radiomics model to predict MVI status after dimensionality reduction using the LASSO method. The AUCs of the combined radiomics model, including age and alpha-fetoprotein levels, in the training, testing, and independent validation cohorts were 0.806, 0.803, and 0.796, respectively. These results indicate that CT and MRI exhibit comparable predictive performance for MVI in solitary HCC. Furthermore, Meng et al[33] demonstrated that MRI radiomics analysis outperformed CT in predicting MVI for HCC tumors larger than 2 cm and up to 5 cm. Despite the reported advancements, various studies employ differing methods for the extraction of texture features or radiomics features, as well as for dimensionality reduction. Additionally, the lack of complete standardization in imaging presents a significant challenge. Furthermore, the variability in combined radiological signs and laboratory indicators is a critical factor contributing to the biases observed in the results. For HCC, dynamic contrast-enhanced MRI of the liver has become the preferred technology for clinical detection, diagnosis, staging and efficacy evaluation. It has advantages over dynamic contrast-enhanced CT in evaluating whether the portal vein, main hepatic vein and its branches are invaded. Therefore, this study mainly focuses on multiple Modal MRI studies. The ANN model can better analyze complex nonlinear relationships than traditional radiomics models (such as support vector machines, random forests, etc.), and is widely used in disease diagnosis, classification, prediction, and survival analysis[41,42]. In this study, the ANN model constructed by combining the dominant features of the group (pseudocapsule + tumor blood vessels + AP + VP) has higher prediction performance for MVI status. This model is an automatic classification prediction with strong operability and stability. It is good and provides a theoretical basis for clinical prediction of HCC recurrence and metastasis. It has great clinical application value. Therefore, it can be recommended as a non-invasive method to predict the MVI status of HCC before surgery.

Clinical application of ANN model for HCC

Previous retrospective study had developed an ANN model to predict post-hepatectomy early recurrence in HCC patients without macroscopic vascular invasion and achieved satisfactory discriminatory and calibration capacities in both the derivation and validation cohorts with greater prediction capacity than a Cox proportional hazards model, some preexisting recurrence models, and commonly used staging systems[42]. Other studies have shown that ANN diagnostic blueprint established by feature genes showcased robust and transferrable prognostic potentialities, portending to alleviate patient encumbrance and elevate life quality[43], and that the ANN model has better diagnostic capabilities than other commonly used models and scoring systems in assessing liver cirrhosis risk in patients with hepatitis B virus-related HCC[44]. Moreover, it was reported that the ANN model was more accurate in predicting 5-year mortality compared to the conventional logistic regression model for HCC[45]. In our present study, we found that in the ANN models based on the relevant radiological signs can improve the performance of texture features in predicting HCC MVI, which can be recommended as a non-invasive method for preoperative prediction of HCC MVI status. have been designed to attain superior classification accuracy through automated extraction of features from images, as well as to enhance the precision in forecasting biological traits and outcomes, including MVI and the likelihood of tumor recurrence[46]. Nonetheless, significant obstacles, particularly regarding interpretability, have impeded their application in clinical environments. To facilitate the effective integration of these models into clinical practice, additional studies are essential to confirm their efficacy and improve their interpretability. Extensive, large-scale, multi-center validation is necessary prior to real-world clinical use.

Limitations and future directions

(1) This is a retrospective, single-center study with a relatively small sample size. Many patients who were clinically considered “high risk” but did not receive surgical treatment were excluded, which may lead to potential selection bias affecting the repeatability and comparability of the results. Further external validation and prospective data validation are thus needed; (2) The construction of the ANN models in this study was completed using MaZda 4.6 in a one-stop manner, with a single reader involved, rendering it impossible to assess the consistency between different readers; and (3) More comprehensive clinical data such as relevant laboratory indicators were not included in the analysis, and the interpretability of the model was not discussed. This is also the direction for further exploration.

CONCLUSION

In summary, in the ANN models of this study, the relevant radiological signs can improve the performance of texture features in predicting HCC MVI to some extent. The ANN model constructed on texture features of multiparametric MRI combined with radiological signs can be recommended as a non-invasive method for preoperative prediction of HCC MVI status. This model is automated for classification prediction, highly operable, and stable. Our findings also suggest that current computer technology applications in disease diagnosis and treatment should not overly rely on the prediction results obtained solely from primary lesion image information by artificial intelligence.