Preoperative prediction of hepatocellular carcinoma microvascular invasion based on magnetic resonance imaging feature extraction artificial neural network

Abstract

BACKGROUND

Hepatocellular carcinoma (HCC) recurrence is highly correlated with increased mortality. Microvascular invasion (MVI) is indicative of aggressive tumor biology in HCC.

AIM

To construct an artificial neural network (ANN) capable of accurately predicting MVI presence in HCC using magnetic resonance imaging.

METHODS

This study included 255 patients with HCC with tumors < 3 cm. Radiologists annotated the tumors on the T1-weighted plain MR images. Subsequently, a three-layer ANN was constructed using image features as inputs to predict MVI status in patients with HCC. Postoperative pathological examination is considered the gold standard for determining MVI. Receiver operating characteristic analysis was used to evaluate the effectiveness of the algorithm.

RESULTS

Using the bagging strategy to vote for 50 classifier classification results, a prediction model yielded an area under the curve (AUC) of 0.79. Moreover, correlation analysis revealed that alpha-fetoprotein values and tumor volume were not significantly correlated with the occurrence of MVI, whereas tumor sphericity was significantly correlated with MVI (P < 0.01).

CONCLUSION

Analysis of variable correlations regarding MVI in tumors with diameters < 3 cm should prioritize tumor sphericity. The ANN model demonstrated strong predictive MVI for patients with HCC (AUC = 0.79).

Key Words: Hepatocellular carcinoma; Microvascular invasion; Artificial neural network; Magnetic resonance imaging; Tumor sphericity; Area under the curve

Core Tip: Constructing an artificial neural network (ANN) using magnetic resonance imaging can accurately predict the presence of microvascular invasion (MVI) in patients with hepatocellular carcinoma (HCC) with tumors < 3 cm. Tumor sphericity has emerged as a significant factor associated with MVI occurrence, emphasizing its importance in MVI analysis. The strong predictive capability of the ANN model, with an area under the curve of 0.79, shows its potential to enhance MVI prediction accuracy and aid in the management of patients with HCC.

INTRODUCTION

Hepatocellular carcinoma (HCC) is the fifth and seventh most prevalent tumor in men and women, respectively. However, its incidence is rising continuously[1]. Surgical resection remains the most effective treatment for HCC. Advances in imaging technology have augmented the detection rate of early-stage HCC, consequently enhancing overall survival rates[2,3]. Nevertheless, the persistently high recurrence rate after surgical intervention continues to significantly affect the prognosis of patients with HCC[4]. Microvascular invasion (MVI) in HCC denotes the presence of a cluster of cancer cells within the lumen of endothelial cell-lined vessels, as observed by pathologists under a microscope. MVI is commonly considered positive when the number of cancer cells suspended in the vascular lumen exceeds 50[5]. The presence of MVI in HCC is widely acknowledged to indicate a higher likelihood of tumor spread to adjacent tissues, signifying a more aggressive disease phenotype. MVI is a significant risk factor for poor disease-free survival and overall survival in patients with HCC[6]. A previous study highlighted the strong correlation between MVI and HCC recurrence within the first two years of surgical treatment[7]. It is reported a 5-year disease-specific survival rate of 59.3% for patients with MVI compared to 92.0% for those without MVI after curative resection for HCC[8,9]. Additionally, in a postoperative investigation involving 108 patients (32 with MVI and 76 without MVI), Liu et al[10] demonstrated that a wide surgical margin (> 10 mm) was associated with improved recurrence-free survival in patients with HCC and MVI presence. These findings underscore the importance of tailoring surgical approaches based on MVI status to improve patient prognosis.

However, the diagnosis of MVI in HCC currently relies solely on the pathological examination of the HCC tissue acquired during surgery. Unfortunately, radiologists cannot directly identify MVI by visual inspection using computed tomography (CT) or magnetic resonance imaging (MRI). Consequently, the strategy for treating patients with HCC often necessitates the commencement of liver resection, limiting the diagnostic value of MVI. Tumor biopsy is suggested for diagnosing MVI; however, its clinical utility has not been validated for several reasons. First, the amount of tissue obtained via biopsy is typically insufficient to ensure an accurate diagnosis. Second, tumor biopsy poses the risk of tumor implantation along the needle tract[11].

Therefore, demand is evident concerning a noninvasive preoperative radiological method to determine MVI in HCC, which can facilitate informed personalized treatment planning for patients. However, the complexity and nonlinearity of the information contained in MRI images pose challenges for quantification. Therefore, distinguishing between feature categories is difficult using a traditional linear discriminant analysis. However, an artificial neural network (ANN) has demonstrated the ability to outperform traditional discriminant analysis methods in calculating these features[12].

Currently, studies on MVI rely predominantly on traditional statistical methods that require additional imaging sequences. To the best of our knowledge, a notable gap exists in studies focusing on patients with small HCC tumors measuring < 3 cm in diameter, particularly in terms of feature analysis and MVI prediction based on MRI T1 plain scans[13]. This retrospective study aimed to address this gap by screening features from MRI T1 plain scan images that were significantly associated with MVI in HCC tumors < 3 cm in diameter. We constructed an ANN model based on MRI texture features and evaluated its performance in predicting the presence of MVI in HCC.

MATERIALS AND METHODS

Study population

Between October 2016 and October 2022, data from patients meeting the following inclusion criteria were collected at Beijing Tsinghua Changgung Hospital: (1) Diagnosed with HCC; (2) Presence of a single lesion meeting the Milan criteria; (3) Lesion diameter < 3 cm; (4) Underwent MRI examinations within 1 month prior to surgery; (5) No prior treatment before the MRI examination; and (6) No history of other malignancies. Further exclusions included the following: (1) Patients who did not undergo hepatectomy after MRI examination; (2) Poor-quality MR images; and (3) Patients with contraindications related to MR examination (such as metal implants and claustrophobia). A total of 255 patients were included in this study, and the screening process is illustrated in Figure 1.

Figure 1 Flowchart of the search strategy used according to research requirements. HCC: Hepatocellular carcinoma; MVI: Microvascular invasion.

Among the included patients, 198 (77.6%) were men. The mean patient age was 52 years. Of the 255 patients diagnosed with HCC, 150 exhibited MVI, representing a ratio of 58.8%. The additional clinical parameters of these patients are presented in Table 1.

Table 1 Clinical information of the 255 hepatocellular carcinoma patients in the study.

Variables	MVI present (n = 150)	MVI absent (n = 105)	P value
Age	52.1	52.0	0.8551
Male sex, n (%)	118 (78.70)	80 (76.20)	0.3822
Basic condition			0.1712
Hepatitis C	5	6
Hepatitis B	132	86
Alcoholic liver	1	1
Multiple	12	12
AFP values	193.8	195.9	0.1661
AFP grade (+), n (%)3	68 (45.30)	44 (41.20)	0.5642
Differentiation			0.9882
Low	6	1
Medium-low	1	2
Medium	137	77
Medium-high	3	8
High	3	17
MVI positive, n (%)	111 (58.12)	40 (62.50)	0.5602

¹Statistics are analyzed using Mann-Whitney U test, respectively.

²Statistics are analyzed using χ², respectively.

³When the alpha-fetoprotein (AFP) value is greater than 10, we consider its AFP grade to be positive (+).

MVI: Microvascular invasion.

MRI series

Patients underwent MRI using 1.5T Signa HDx and 3.0T Signa HDxt magnetic resonance scanners (GE HealthCare Technologies, Chicago, IL, United States). The following sequence scan parameters corresponded to 1.5T and 3.0T, respectively: T1-weighted imaging (T1WI) plain LAVA Mask and three-dimensional dynamic enhanced scanning, with enhanced equilibrium phase coronal position: TR 3.0-4.0 ms, TE 1.5-2.0 ms, layer thickness 5.0 mm, layer spacing -2.5 mm; TR 3.0-4.0 ms, TE 1.0-1.5 ms, layer thickness 4 mm, layer spacing -2.0 mm. Notably, T1WI plain scan images were primarily used in our study. Following hepatectomy, the lesion tissue sections were examined microscopically to determine the presence of MVI based on the pathology. MVI was considered present when the number of cancer cells within the vascular lumen was ≥ 50. Subsequently, patients were categorized based on their MCI status, with those without MVI marked as 0 and those with MVI labeled as 1. All 255 patients were assessed by radiologists, and their original images were registered to facilitate feature extraction.

Feature collection

Radiomics entails the conversion of medical images by delineating regions of interest into high-dimensional characteristic data for analysis using computer image-processing techniques[14,15]. Through this transformation, valuable information regarding the diagnosis of MVI can be obtained.

Radiomics can quantify image features, including shapes and textures, which may not be quantifiable using other methods. Additionally, it can reflect biological characteristics such as internal and spatial heterogeneity.

In this study, PyRadiomics (https://pyradiomics.readthedocs.io) was used for the high-throughput extraction of quantitative features. PyRadiomics is an open-source software program based on Python specifically designed for extracting radiological data from medical images[16]. Our research employed the PyRadiomics algorithm to facilitate the extraction of radiomics features. First, image loading and preprocessing tasks, such as resampling and clipping, were performed using SimpleITK. Subsequently, the loaded data were converted into a Numpy array for further evaluation using multiple classes of elements. In addition, an optional filter was integrated into the process. Notably, PyRadiomics has been successfully applied to the study of various cancers[17].

In our study, PyRadiomics facilitated the extraction of 110 dimensional features from MR images, which were classified into seven categories: First-order statistics, shape-based, gray-level co-occurrence matrix, gray-level run length matrix, gray-level size zone matrix, neighboring gray-tone difference matrix, and gray-level dependence matrix.

Analysis

For the extracted features, data analysis initially involved identifying abnormal values, followed by normalization of features that met specific criteria to facilitate the training of the ANN. A bivariate Spearman test was conducted for each feature with respect to MVI to calculate the correlation and corresponding p-values. Features with P values < 0.05 were deemed significantly correlated with the MVI classification. Following this procedure on the 110-dimensional feature set, we found that 47-dimensional features exhibited a significant correlation with MVI. The screening process for the entire feature set using methods 2-2 to 2-4 can be referred to in Figure 2.

Figure 2 Tumor contours were labeled on each patient’s T1-weighted imaging plain image. After preprocessing, image features were acquired by pyradiomics, and then the correlation between features and microvascular invasion. was analyzed. Features with P value < 0.05 were selected as input to the artificial neural network. GLCM: Grey level co-occurrence matrix; GLRLM: Grey level run length matrix; GLSZM: Gray level size zone matrix.

Model training method

Variable features considered to be significantly correlated with the MVI in the above analysis were regarded as the input of the ANN. In the data assignment process, 5-fold cross-validation was performed. Each fold set consisted of 255 patients randomly assigned to two groups: Group A (for training, 191 patients, 75%) and Group B (for testing, 64 patients, 25%). To enhance the reliability of the results, training was performed five times, and the average area under the curve (AUC) was calculated. Moreover, an ANN comprising three hidden layers was established. Each classifier contained 47 random input neurons, and the number of neurons in each hidden layer ranged from 20 to 50 (Figure 3). This approach effectively increases the randomness of individual classifiers by biasing them toward different features. In this study, the learning rule employed was the back-propagation of errors, allowing for the adjustment of internal network parameters during repeated training cycles to reduce the overall error. Training was terminated when the sum of the squared errors reached a minimum. Subsequently, the trained neural network model was verified using a test group. Sensitivity, specificity, and the AUC were calculated and obtained as evaluation metrics.

Figure 3 The basic structure of the artificial neural network constructed in this study. MVI: Microvascular invasion.

The bagging strategy is a common algorithm used in ensemble learning that combines the predictions of multiple weak classifiers without strong dependencies. This method can enhance classifier classification results and significantly improve the generalization ability and accuracy of the model. In this study, 50 weak classifiers were trained using the ANN described in Section 2.5. Each classifier iterates the input data 1000 times to minimize the loss function. Random numbers were introduced to determine the number of training sets and neurons in each hidden layer, thereby enhancing the randomness of each classifier. Subsequently, the bagging process involves voting among the weak classifiers. If more than half of the classifier results indicated MVI (i.e., the patient was judged to have MVI), the bagging result was classified as 1; otherwise, it was classified as 0.

RESULTS

After applying the bagging technique, the AUC reached 0.79, whereas the AUC values of the top three neural network models reached 0.74, 0.73, and 0.71, respectively. Evidently, employing the bagging technique increased AUC values.

Figure 4 illustrates the results obtained after applying the bagging strategy, including metrics such as the AUC. The top three classifiers with the highest accuracies are shown in the figure.

Figure 4 The results of the bagging strategy. The red line represents achievement of area under the curve (AUC) of 0.79. The top three classifiers out of the 50 classifiers are represented by blue, green, and orange lines, corresponding to AUC values of 0.74, 0.73, and 0.71, respectively. ROC: Receiver operating characteristic.

Table 2 presents the correlations between the various indicators and the MVI. Surprisingly, tumor volume, alpha-fetoprotein (AFP) levels, and AFP classification did not correlate with MVI, contrary to the findings of previous studies. However, tumor non-roundness, calculated using the formula, where V represents tumor volume and A represents tumor surface area obtained through definite integration of the tumor-labeled mask, showed a significant correlation with MVI (P < 0.01). A non-roundness value of 1 indicates standard sphericity, with smaller values indicating greater tumor non-roundness. Moreover, non-roundness was the only relevant variable among the 16-dimensional features extracted from tumor shape. Therefore, in predicting MVI in tumors < 3 cm in diameter, more attention should be focused on tumor non-roundness than on tumor volume.

Table 2 Correlation analysis between related features and microvascular invasion.

Variables	Mean	P value
Male, n (%)	198 (77.6)	0.001a
Age	52	0.619
AFP values	194	0.606
AFP grades, n (%)	112 (43.9)	0.495
Original_shape_MeshVolume	6819.940 (mm3)	0.104
Original_shape_Sphericity	0.734	0.008
Original_shape_MajorAxisLength	25.33151 (mm)	0.113

^aP < 0.001.

AFP: Alpha-fetoprotein.

DISCUSSION

HCC is the third leading cause of cancer-related mortality worldwide. The global incidence of HCC is 16 cases per 100000 individuals. Surgical resection remains the most effective treatment option for patients with HCC and has the best prognosis. However, even after successful resection, patients continue to face a high risk of postoperative HCC recurrence[18].

MVI is a powerful tool for predicting HCC recurrence after hepatic resection and transplantation[19-21]. However, MVI can only be confirmed through microscopic examination because its diagnosis surpasses the resolution of conventional imaging data. The incidence of MVI varies significantly among patients, ranging from 15.0%-57.1%. The ability of MVI to infiltrate into other branches of the portal venous system can alter tumor perfusion, thereby increasing the likelihood of metastasis and reducing progression-free survival[22,23].

Therefore, the development of a feasible, non-invasive method to diagnose MVI preoperatively is imperative, enabling surgeons to determine optimal treatment strategies for patients with HCC[24]. However, a unified clinical method for predicting MVI before surgery currently remains unavailable.

As a non-invasive preoperative examination, a computer-aided MVI diagnostic algorithm can mitigate the risks associated with invasive biopsy procedures. Moreover, it relieves psychological stress and aids doctors in formulating targeted surgical plans.

Currently, various methods are available for predicting MVI in published studies. These methods include investigating non-smooth margins on multiphasic CT images[25], observing the correlation between MVI and peritumoral hyperintensity observed on hepatobiliary phase images of gadoxetate disodium-enhanced MRI[26,27], assessing MVI in patients with small HCC using liver CT perfusion[28], and measuring apparent diffusion coefficients for lesions[29].

However, determining whether a tumor margin is smooth often relies on manual judgment, making the criteria subjective. In this study, we integrated a mask matrix with tumor labeling to obtain the volume and surface area of the tumor. Subsequently, the out-of-roundness was calculated using these parameters. This method offers greater objectivity and quantifiability in data analysis.

Moreover, gadoxetate disodium-enhanced MRI and perfusion liver CT require additional sequences, leading to extended examination times. Furthermore, the processing of such data requires complex algorithms. In contrast, our study used routine MRI T1 plain scan sequences without supplementary specialized sequences.

Jonas et al[19] reported a correlation among tumor diameter, nodule number, and histopathological classification, albeit with limited predictability for HCC vascular invasion exceeding 5 cm. Currently, studies of MVI predominantly use traditional statistical methods and require additional imaging sequences. To the best of our knowledge, studies focusing on patients with HCC with tumors < 3 cm, as well as on feature analysis and MVI prediction based solely on MRI T1 plain scan, remain lacking.

Despite the merits of this study, it also had a few limitations. This study gathered data from a single center (Beijing Tsinghua Changgung Hospital), which may affect the robustness of the constructed ANN. However, the ANN was superior to other linear models in the diagnosis of MVI. Moreover, the inherent learning ability of neural networks suggests that their performance can be enhanced by incorporating new data from more patients with HCC.

CONCLUSION

In summary, this study employed a radiomics model to extract and quantify 110-dimensional features from T1 plain MRI scans of 255 patients. Following feature analysis and screening, 47-dimensional features were identified and retained as inputs for the ANN. Subsequently, an ANN model was established to predict the MVI, achieving an AUC of 0.79, thereby exhibiting strong generalization ability owing to the bagging strategy.