Abstract
Microvascular invasion of hepatocellular carcinoma is an important factor affecting tumor recurrence after liver resection and liver transplantation. There are many ways to classify microvascular invasion, however, an international consensus is urgently needed. Recently, artificial intelligence has emerged as an important tool for improving the clinical management of hepatocellular carcinoma. Many studies about microvascular invasion currently focus on preoperative and prognosis prediction of microvascular invasion using artificial intelligence. In this paper, we review the definition and staging of microvascular invasion, especially the diagnosis of it by using artificial intelligence. In preoperative prediction, deep learning based on multimodal data modeling of radiomics-screened features, clinical features, and medical images is currently the most effective means. In prognostic prediction, pathology is the gold standard, and the techniques used should more effectively utilize the global features of the pathology images.
Introduction
Hepatocellular carcinoma (HCC) is one of the most common malignancies in the world. This tumor tends to invade intrahepatic vessels, leading to intrahepatic and extrahepatic metastases 1 and a poor prognosis. The most current treatment options for specific HCC patients are hepatectomy and liver transplantation. However, due to the high incidence of tumor recurrence and metastasis, the tumor recurrence rate is 70% at 5 years after liver resection and 15-30% after liver transplantation, and the long-term survival outcome remains unsatisfactory.2,3
Microvascular invasion (MVI) is a manifestation of the aggressive biological behavior of the tumor and is currently one of the most critical factors in predicting HCC recurrence. 4 MVI increases the risk of postoperative recurrence and metastasis in patients with HCC and is a significant factor affecting long-term prognosis. Therefore, early assessment of the presence of MVI and its progression and rational individualized treatment has become a hot spot in recent years in liver cancer research. In some cases, there may be a lack of distinction between macrovascular invasion and MVI, with no agreed criteria or terminology to distinguish between MVI of different severity. 5 Therefore, it is particularly important to predict MVI preoperatively and to have a consistent and accurate international consensus on MVI.
Since MVI is at the microscopic level, it cannot be judged preoperatively by the physician's naked eye alone on the imaging data and can consume a lot of the pathologist's time postoperatively. In contrast, artificial intelligence (AI) can combine image features with clinical data in a statistical-like manner to make a diagnosis of MVI. Imaging features such as tumor size, unsmooth tumor margins, and peritumoral enhancement are markers that have been reported by AI to have some correlation with MVI.6,7 Deep learning is now widely used in medical image processing, 8 and its ability to discriminate benign and malignant diseases by medical images has surpassed that of senior imaging experts and pathologists.9,10 Radiomics, which has emerged in recent years, can extract and process a large number of features from medical radiology images through data characterization algorithms to aid in disease diagnosis and predict treatment response.11–13 The relationship between them can be seen in Figure 1. In addition, the deep learning method usually needs time-consuming segmentation during training but may become automatic during inference. The radiomics methods are usually based on feature extraction during training but may need manual segmentation during inference.
The relationship between artificial intelligence (AI) and some of its subcategories.
Previous reviews have almost exclusively focused on discussing the preoperative prediction of MVI using imaging, and there are some limitations in the discussion of MVI study reports. Therefore, the purpose of this paper is to introduce MVI and to summarize AI studies on patients with HCC with MVI in terms of preoperative prediction of MVI status and postoperative prediction of long-term patient survival. Discussions of blood tests, ultrasound images, and pathologic methods are added to the clinical data, computed tomography (CT), and magnetic resonance imaging (MRI) to provide a more comprehensive overview of currently reported studies. It helps and guides clinicians and researchers in understanding and treating MVI more comprehensively from multiple directions.
Definition and Classification of MVI
MVI is a pathological concept. It is defined in China as a microscopic representation of MVI as a nesting mass of cancer cells in the lumen of a vessel lined by endothelial cells. 14 The histological types of MVI include small venous thrombi, small arterial thrombi, small bile duct thrombi, pericyclic vascular thrombi, lymphatic duct thrombi, and suspended cancer cells.
In this paper, 6 representative MVI classification studies are listed in Table 1. Roayaie et al 15 concluded that invasion of a vessel with a muscular wall and invasion of a vessel that is more than 1 cm from the tumor can accurately predict the risk of recurrence and survival of patients with MVI after resection of HCC. Patients with vascular invasion were also classified into 5 groups according to the risk factors classified. This classification was widely accepted by investigators in the first few years, but treatment recommendations for patients with MVI with poor prognosis were lacking. MVI grades were classified according to risk factors, and poorer MVI scores were associated with larger tumor size, elevated serum alpha-fetoprotein (AFP), and high tumor grade, as has been shown in other studies. 16
Pathological Classification Pattern of Some MVI.
Abbreviations: NA, not available; HCC, hepatocellular carcinoma; MVI, microvascular invasion.
Sumie et al 17 devised a new way of classifying MVI based only on the number of invaded vessels. Due to the many factors affecting MVI, this classification method considering only the invaded vessels is relatively imperfect and may not work well with additional external validation. Iguchi et al 18 found that the number of infiltrating cells in the vascular lumen correlated with HCC prognosis, and the number of cancer cells observed in the vascular lumen greater than 50 was used as a threshold value to classify MVI. Multivariate analysis showed that both mild and severe were independent predictors of poor prognosis, especially severe MVI was an important predictor of recurrence-free survival. However, the staging model was only applicable to patients with HCC who met the Milan criteria. Feng et al 19 focused on the morphological characteristics of intravascular tumor invasion while combining the number of observable MVI vessels to establish 4 MVI levels. This classification has been reported to improve the ability in predicting early recurrence in patients with MVI. However, all these MVI classifications are based only on the characteristics of MVI, without considering any clinicopathological features of patients. In a recent study, Wang et al 6 classified MVI by building a scoring model using score intervals. The scoring system was modeled using AI, and input data included patient survival time, maximum tumor diameter, MVI area, and cirrhosis. The classification system was reported to have better predictive power than any other staging system. It also shows that AI has some advantages in predicting MVI. These clinical MVI classification systems predicted more accurately postoperative long-term overall survival (OS) for HCC patients with MVI after R0 liver resection and can help physicians make targeted classified treatments after a patient's surgery by grouping patients together.
In studies of MVI classification, there are differences in the methods used and in the conditions of case screening that may affect the results. However, the number of invading cancer cells, the number of invading vessels, and the distance from the tumor are currently the more accepted assessments. The new scoring system proposed by Wang et al 6 combines previous studies more comprehensively with better results.
Artificial Intelligence Preoperative Diagnosis of MVI
In recent years, the development of imaging techniques such as CT and MR, as well as the characterization of new serum biomarkers and biopsied tissues, have provided more reliable methods for AI to predict MVI status preoperatively. 20 These AI prediction methods are divided into 2 main categories: imaging histology and deep learning. At present, the existing deep neural network-based automatic diagnosis method of MVI is mainly based on direct training, and the common method is to use CT and MR multimodal data for feature fusion. Some of the preoperative prediction models based on CT, MR, and ultrasound are listed in Tables 2 to 4 (clinical data and blood tests not tabulated).
AI Models for Preoperative Prediction of MVI Based Partially on CT.
Abbreviations: NA, not available; CNN, convolutional neural network; AUC, Area Under Curve; val, Validation; MVI, microvascular invasion; AI, artificial intelligence; CT, computed tomography.
Data Enhancement.
AI Models for Preoperative Prediction of MVI Based Partially on MRI.
Abbreviations: NA, not available; CNN, convolutional neural network; AUC, Area Under Curve; val, Validation; MVI, microvascular invasion; AI, artificial intelligence; CT, computed tomography; MRI, magnetic resonance imaging.
The use of Gd-EOB-DTPA enhanced MRI.
The use of Gd-DTPA.
AI Models for Preoperative Prediction of MVI Based Partially on Ultrasound.
Abbreviations: NA, not available; AUC, Area Under Curve; val, Validation; CNN, convolutional neural network; MVI, microvascular invasion; AI, artificial intelligence.
Of the 25 models included in this paper, 19 were built on single-center, retrospective studies that used internal validation methods (randomized split or cross-validation) to assess the performance of the MVI prediction model. Three models were based on multicenter retrospective studies and 3 were built on multi/single-center prospective studies that used external validation methods. All patients were diagnosed with HCC based on postoperative pathological specimens and had available preoperative imaging including CT, MR, or ultrasound.
Artificial Intelligence Model Based on Clinical Data and Blood Tests
Artificial neural network is a network structure that can be used to deal with practical problems with multiple nodes and multiple output points and includes various network algorithms such as multilayer perceptron, back-propagation neural network, convolutional neural network (CNN), and recurrent neural network, which are widely used. 45 Its ability to predict HCC classification and MVI status preoperatively by inputting clinical data from patients as well as the number, size, and volume of tumors is superior to the performance of traditional linear models. 46
Serum biomarker-based models have been previously reported for predicting the likelihood of HCC in patients with chronic liver disease and are superior to ultrasound. 47 Chen et al 48 found that preoperative risk prediction for MVI could also be achieved by integrating the results of preoperative blood tests (complete blood count, blood tests, and AFP tests), combined with deep learning techniques. The “Interpretation based Risk Prediction” method was reported to accurately estimate the MVI risk and obtain better performance with a c-index of 0.9052 in the independent validation cohort. In addition, the quantified risk of MVI in the model can be used as an independent preoperative risk factor for recurrence-free survival and OS in HCC patients.
Artificial intelligence models that rely on clinical data and blood test data demonstrate some feasibility; however, more studies will combine CT images and MR images on top of this to obtain more features and achieve better results.
Artificial Intelligence Model Based on CT Images
CT images have been widely used in clinical practice due to their special diagnostic value. Arterial phase (AP), portal venous phase (PVP), and delayed phase (DP) of CT are mainly used for feature extraction and modeling in combination with clinical factors.
It has been suggested that a radiomics model combining clinical factors and CT PVP radiological features has the best predictive effect. 21 Moreover, image features extracted from the intratumoral and peritumoral regions of the PVP have more significant predictive performance and are similar to multidisciplinary Team-like radiomics fusion models, which outperform independent prediction models. 25 However, it has also been shown the better performance of models based on AP and DP. 26 This occurrence may be related to the inclusion criteria of radiomic features. In addition, radiomics features extracted from the core tumor region in combination with Zone of Transition detection can also enhance the prediction. 27
Xiao et al 28 predicted the MVI using the 3DResNet-18 network based on a multicenter retrospective study. The authors proposed a skeleton-sharing model that implements the prediction model into 2 other cohorts, which can significantly improve the performance of the original model. In addition, the introduction of expert knowledge during the training process also improves the predictive ability of the model,28,29 and the AUC of Xiao et al 28 improves from 0.54 to 0.83.
The recently reported MVI prediction model based on the transformer model showed informative results with an AUC score of 0.935 and an accuracy of 0.972. 30 Specifically, operating under the Vision Transformer model proposed by Dosovitskiy et al, 49 the input image is cut into a series of patches, and a sequence of linear embeddings of these patches is provided as input to the transformer, which is processed using a standard transformer. The algorithm of transformer differs from convolution in that it calculates the magnitude of the attentional weights for each pixel location on the image, allowing the network to select the target region that requires more attention. Thus, the unique architecture allows for powerful modeling of global environments and feature representation, and it has been applied to the field of computer vision to model long-distance dependencies. In addition, the ResNet-18 network AUC is better than ResNet-50 and ResNet-101 in the CNN model control group of this study (0.980, 0.978, and 0.947).
Both radiomics and deep learning approaches have some potential for the preoperative prediction of MVI. To compare the differences between the 2 methods, Jiang et al 23 built 3D-CNN model (DL) and XGBoost model (Radiomics), respectively, and found the deep learning model (AUC = 0.906) to be slightly more accurate than radiomics (AUC = 0.887) in a comparison based on the same dataset. However, disregarding the differences between the 2 methods and combining their strengths, using CNN to extract features on CT images and then combining them with features extracted from radiomics to build a deep learning model can effectively improve the prediction ability.24,50 It can increase the AUC to about 0.9.
Most preoperative studies predicting MVI have focused on predicting its status, and few studies have classified the status of MVI. Zhang et al 22 proposed to focus on multilevel stratification of MVI using radiomics. However, due to the microscopic nature of MVI, there is some difficulty in categorizing MVI status at the imaging level. Moreover, the importance of clinical staging to improve the prediction of MVI status has been demonstrated in studies. 51 However, clinical staging may be associated with specific radiomic features, which leads to the failure of clinical staging to improve the predictive validity of MVI status models in some studies.
Artificial Intelligence Model Based on MR Images
Compared to CT, MR is capable of multimodal and multidirectional assessment of lesions and can better characterize soft tissue features, atomic signal intensity, and lesion enhancement, providing more anatomical and functional information about the tumor.
Lee et al 7 demonstrated in a single-center retrospective study that a combination of 2 or more peritumor arterial enhancement, unsmooth tumor margins, and peritumor hepatobiliary phase low signal can be used as a preoperative imaging biomarker for predicting MVI with >90% specificity and is associated with early recurrence after curative resection of a single HCC. Hepatobiliary phase TI-weighted images combined with fusion radiomic features of serum AFP levels, nonsmooth tumor margins, and perianeurysmal enhancement have high accuracy for preoperative prediction of MVI. 31 In addition, extracting features from Gd-EOB-DTPA-enhanced MRI and 3 Gd-DTPA-enhanced MRI sequences and combining intratumor peritumor and different sequences can improve the prediction of preoperative MVI.32–34 Combined with clinical data, the results can be further improved. 35
T2WI and DWI sequences have also been used for tumor feature extraction, and most HCCs exhibit high signal on high b-value DWI. The apparent diffusion coefficient calculated using a monoexponential model based on DWI technique has been reported to be an important risk factor for preoperative prediction of MVI in HCC. 52 Moreover, Intravoxel incoherent motion diffusion-weighted imaging with a range of different b-value images has recently been shown to have good preoperative prediction for MVI. 38
Deep supervision is a technique that adds an auxiliary classifier to some intermediate hidden layers of the deep neural grid as a kind of grid support to supervise the main grid, which is used to solve the problems of disappearing training gradients and slow convergence of the deep neural grid. 53 A deep supervised network based on contrast-enhanced MRI arterial, portal vein, and DP image data reported an AUC of 0.926 for preoperative prediction of the MVI model. 36
Wei et al 37 concluded that Gd-EOB-MRI is a superior modality for MVI assessment than contrast-enhanced CT because Gd-EOB-MRI can better capture these perfusion and functional changes, making predictions more sensitive and accurate. It has been reported to show that combining clinical data, imaging data, and features of CT and MR images from different scanning periods into multimodal data as network inputs for deep learning can more accurately predict MVI preoperatively. 39 Model AUC improved from 0.79 to 0.84. In addition, the features on the CT and MR images in this study were also extracted by the ResNet18 network. Without considering the differences in prediction due to the selection of CT and MR as data sets, the size of the data set and the depth of the network are also key factors affecting the accuracy of the model. For different data set, different depths of the network can have some impact on the results due to overfitting and too many parameters. Therefore, the amount of data and the depth of the network are also important concerns. The study by Cao et al 29 is a typical example that ResNet-101, which performs well in classifying ImageNet data with a large amount of data, instead performs weaker than ResNet-50 and ResNet-18 in classifying MVI in liver cancer images.
Artificial Intelligence Model Based on Ultrasound Images
Ultrasound imaging is operator-dependent and its imaging technique differs from CT and MR. Some researchers using ultrasound techniques to predict MVI believe that CT and MR still have limitations, such as the potential risk of radiation exposure with CT and the relatively expensive and time-consuming nature of MR, and the imaging features are all visible to the naked eye. 40 Grayscale ultrasound is the most commonly used first-line imaging method for preoperative liver cancer and has the unique advantages of being radiation-free, easy to perform, and inexpensive. In 2012, Streba et al 54 prospectively studied contrast-enhanced ultrasound images of 112 patients with HCC, and the accuracy of artificial neural networks was 94.5% in the training and 87.1% in the testing set. It indicates that ultrasonography can be used for preoperative predictive imaging analysis of MVI.
Predictive models based on grayscale ultrasound have shown some effectiveness in extracting radiographic features in the tumor and peritumor regions. 41 It is possible to predict MVI status and classify MVI positivity into M1 and M2. Model performance can be further enhanced by more advanced imaging techniques, and radiological models based on Kupffer phase ultrasound images of tissue near liver cancer lesions show better predictive value. 43
Studies based on the preoperative prediction of MVI from ultrasound images are further developing, and the usefulness of ultrasound images in predicting MVI was also confirmed by the retrospective studies of Zhang et al 42 and Hu et al 40 whose established contrast-enhanced ultrasound image radiomics nomogram had better predictive performance than the clinical data model, and multimodal enhanced ultrasound images were better than grayscale images. Deep learning methods have also been reported in ultrasound images, 55 and in addition, a deep learning–based study reported on MVI prediction metrics shows the feasibility of microscopic features of ultrasound images in this field. 44
Artificial Intelligence–Assisted Prognosis Prediction for MVI Patients
As mentioned previously, MVI is an independent risk factor affecting patient prognosis; therefore, discussing the different risk levels as well as survival and recurrence risks of patients with MVI is a more meaningful task in the absence of a clear definition of MVI.
There are currently 3 main clinical data used to analyze patient prognosis, clinical data, imaging data, and whole slide image (WSI) of postoperative pathology. Artificial intelligence–assisted prognostic studies are mainly based on combining clinical data with imaging and WSI, respectively, and then risk scoring corresponding to the follow-up data of patients.
Among the previous studies that did not use postoperative pathology sections of HCC patients, the more typical ones are Zhang et al 44 based on ultrasound images, Chen et al 47 based on blood tests, Liu et al 38 based on MR images, and Zhou et al 28 based on CT images to predict the occurrence of MVI preoperatively and correlate prognosis of HCC patients, pointing out that MVI is an independent risk factor for patient prognosis, and worse MVI status corresponds to worse prognosis, which has been reported similarly in many articles.
Among the studies based on pathological images, Shi et al 56 studied the prognosis and classification of HCC, and the problem of excessive memory of the whole WSI pixels was solved by cutting the WSI into smaller patches by the traditional patch-cutting method and using each small patch as input data. Yamashita et al 57 to provide a risk score for disease recurrence after primary resection may enhance current patient stratification methods and help improve clinical management of patients undergoing stage I surgical resection for HCC. Chen et al 58 explored factors affecting MVI status from WSI with abundant blood sinus, abundant tumor mesenchyme, and highly intratumor heterogeneous trabecular structures were identified as key features associated with MVI-positive, whereas severe immune infiltration and highly differentiated tumor cells were associated with MVI-negative. Wang et al 18 included MVI area as a prognostic factor in a predictive model, which did not utilize pathological image features. Rather, the area of MVI was calculated on the WSI, and risk assessment was performed using a simple 2-layer fully linked neural network and text data.
Pathology is the gold standard for diagnosing the benignity and malignancy of tumors, and the information on its postoperative sections is also extremely important for patient prognosis determination. However, conventional pathology usually cuts WSI into many small patches during AI analysis, and this approach results in the weights learned by the network only targeting the information on each small patch, thus ignoring the contextual environment in the WSI. Therefore, there is still room for improvement in the choice of network. Graph neural networks have been proven to be effective in WSI analysis59,60 and may be considered for use in the prognostic analysis of patients with liver cancer.
Summary and Prospect
Several retrospective trials have shown that the presence of MVI is a key determinant of recurrence after surgical resection and transplantation. The presence of MVI is highly correlated with poor biologic tumor behavior in HCC and is a more accurate predictor of long-term survival than traditional staging criteria. 5 With the increasing number of future HCC patients and advances in scientific research technology, there is an urgent need to reach an international consensus on the definition and classification of MVI in HCC to curative treatment after surgery to provide a more consistent assessment and reliable prediction of surgical outcomes.
In the numerous studies included, it was shown that models combining CT and MR image data in the arterial and portal phases had a good improvement in predictive performance. In the selection of CT and MR images, Zhang et al 61 concluded that the difference in performance between the 2 could be ruled out. In the study by Zhou et al, 36 25 clinical data fluctuated during the DP (1-3 min) and at different locations (coronal), and the contrast agent spilled from the tumor area during the DP. Therefore, the DP is not recommended to produce promising results for MVI prediction. It is noteworthy that disease-free survival and OS were generally lower in MVI-positive patients in patient prognostic assessment models, consistent with clinical judgment.
In contrast to deep learning models, non-deep learning models mostly use LASSO regression algorithms to select features, and logistic regression implements a combination of radiomic features and clinical factors. Classification methods include machine learning algorithms such as SVM, random forest, and XGBoost. These operations are relatively patterned and fixed, and their focus is on selecting the different stages of scanning image slices. The deep learning model, on the other hand, has network model selection in addition to image slice selection due to the autonomous learning characteristics of its deep neural network. The neural network shows advantages in feature selection and prediction results after completing the training. ResNet18 is the most reported network that has been used for modeling predictive MVI, which has shallower layers and seems to have better results for medical images with small amounts of data.27,29,39 In addition, the combination of clinical features, radiomics, and deep learning into multimodal data has been shown to further improve predictive performance.23,49 However, there are still limitations in these studies. Few of the constructed prediction models are practiced clinically due to poor interpretability and generalizability of the models, 24 which needs to be addressed gradually in future studies. Currently, for the general problem of poor interpretability of AI models, there are 2 categories of optimization methods: one is to interpret neural network models after training, such as class activation map-based methods 62 ; the other is to help deep learning models focus on relevant regions with medical logic by introducing prior knowledge and experience in the training process to guide the construction of deep learning models.27,28
It is important to recognize that the help provided by AI is indispensable due to the microscopic and unpredictable nature of MVI demonstrated on images and the indistinguishability of features from HCC features on histopathological images. Improving the accuracy of AI model predictions and the amount of patient data used to predict MVI are challenges that need to be addressed at this time. In addition, any AI-based algorithm needs to be externally validated in an independent dataset, including the replicability of the study results in other medical centers, as the model may be overfitted, leading to an overestimation of its performance. Moreover, current multicenter studies also have some differences in performance, mainly due to the lack of consideration of the differences in data distribution between different centers and the lack of generalization performance of the network. Therefore, while modeling, researchers should consider how to build up an open database platform to assess the predictability of each model and perform external validation.
As mentioned earlier, in the absence of a clear definition of MVI at the present time, discussing the different risk levels of patients with MVI, as well as long-term survival and risk of recurrence, is a more meaningful task. Classifying and defining MVI on pathology images with the help of AI is more in line with the work of pathologists in assessing MVI. Therefore, future research should focus on how to efficiently utilize the information on pathology images of MVI patients to classify MVI by rationally and accurately assessing the risk level of patients. In addition, the establishment of public databases is a top priority.
Conclusion
Artificial intelligence has provided significant help in studies targeting patients with MVI with HCC, and its importance cannot be ignored. In preoperative prediction, deep learning based on multimodal data modeling of radiomics screening features, clinical features, and medical images is currently the most effective tool. In prognostic prediction, pathology is the gold standard, and the AI methods used should more effectively utilize global features on histopathological images, which is a key research direction for future researchers to focus on.
Footnotes
Declaration of Conflicting Interests
The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.
Funding
The author(s) disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: This work was supported by the Key Project of the National Natural Science Foundation of China (81730097), the National Natural Science Foundation of China (82072618), and the National Key Research and Development Program of China (2022YFC2503700).
