Preoperative prediction of hepatocellular carcinoma tumour grade and micro-vascular invasion by means of artificial neural network: A pilot study

Background & Aims

Hepatocellular carcinoma (HCC) prognosis strongly depends upon nuclear grade and the presence of microscopic vascular invasion (MVI). The aim of this study was to develop an artificial neural network (ANN) that is able to predict tumour grade and MVI on the basis of non-invasive variables.

Methods

Clinical, radiological, and histological data from 250 cirrhotic patients resected (n = 200) or transplanted (n = 50) for HCC were analyzed. ANN and logistic regression models were built on a training group of 175 randomly chosen patients and tested on the remaining testing group of 75. Receiver operating characteristics curve (ROC) and k-statistics were used to analyze model accuracy in the prediction of the final histological assessment of tumour grade (G1–G2 vs. G3–G4) and MVI (absent vs. present).

Results

Pathologic examination showed G3–G4 in 69.6% of cases and MVI in 74.4%. Preoperative serum alpha-fetoprotein (AFP), tumour number, size, and volume were related to tumour grade and MVI (p <0.05) and were used for ANN building, whereas, tumour number did not enter into the logistic models. In the training group, ANN area under ROC curves (AUC) for tumour grade and MVI prediction were 0.94 and 0.92, both higher (p <0.001) than those of logistic models (0.85 for both). In the testing group, ANN correctly identified 93.3% of tumour grades (k = 0.81) and 91% of MVI (k = 0.73). Logistic models correctly identified 81% of tumour grades (k = 0.55) and 85% of MVI (k = 0.57).

Conclusion

ANN identifies HCC tumour grades and MVI on the basis of preoperative variables more accurately than the conventional linear model and should be used for tailoring clinical management.

Introduction

Hepatocellular carcinoma (HCC) is one of the five most common malignancies worldwide and often occurs in patients with chronic viral hepatitis and cirrhosis [1]. Although current screening techniques identify approximately one third of the tumours at a potentially curable stage, the high incidence of recurrence, both after surgical and non-surgical approaches, remains a major challenge in HCC management [2], [3], [4], [5]. Tumour histological features strongly influence not only tumour recurrence but also patient survival: nuclear grading and vascular invasion are considered the two most important prognostic factors but they are accurately assessable only with the pathological examination of surgical specimens. Furthermore, only a minority of cirrhotic patients with HCC are eligible for resection or transplantation [3], [4], [5], [6], [7], [8], [9], [10]. Percutaneous fine-needle aspiration (FNA) biopsy can be used to characterize HCC grade but its usefulness remains controversial as the heterogeneity of the histological grade, within a single tumour, may limit overall accuracy [11]; in addition, FNA cannot assess microscopic vascular invasion (MVI). Several authors have investigated the relationship between morphological characteristics of the tumour, its grade, and the presence of MVI. Tumour size and number of lesions, assessed on surgical pathologic specimens, as well as serum alpha-fetoprotein (AFP) are the variables that were most frequently found to be associated with both tumour grade and vascular invasion; given the strict relationship reported between tumour grade and microscopic vascular invasion, histological grade is currently used as a surrogate marker of MVI [11], [12], [13], [14], [15]. However, the interaction among these factors is complex and non-linear [5], [13], [16], thus, making it difficult to distinguish between classes when using the conventional linear discriminant analysis. When classes are separated by a non-linear boundary, as in the present situation, an artificial neural network (ANN) has been demonstrated to perform better than conventional discriminant analysis [17], [18], [19], [20].

The artificial neural network uses computer technology to model a biological neural system both structurally and functionally. Like its biological counterpart, an artificial neural network consists of a set of highly interconnected processing units (neurons) tied together with weighted connections. The network itself consists of an input layer, an output layer, and one or more hidden layers. The input layer comprises the data available for the analysis (e.g. various laboratory tests) and the output layer comprises the outcome (e.g. diagnosis). One of the basic characteristics of the ANN is that it learns through examples: learning is achieved through exposure of paired input–output data (training). An ANN learns to associate each input with the corresponding output, by modifying the weight of the connections between neurons. Once an input has been applied as a stimulus to the first layer of neurons it is propagated through each upper layer until an output is generated. This output pattern is then compared to the desired output and an error signal is generated: the error signal is then transmitted backwards across the net and the connection weight between neurons is updated in order to decrease the overall error of the network. As learning proceeds, the error between ANN output and the desired output decreases until a minimum is reached. Based on the knowledge accumulated during training, the ANN can assign outputs (diagnoses) to new input data not used in the learning process. Thus, after training, the ANN can identify patterns or make predictions on datasets never seen before [17], [18], [19], [20], [21].

The main aim of the present study was to assess the ability of the ANN to predict tumour grade on the basis of preoperative clinical and radiological variables. We also explored its ability to predict MVI and compared ANN performance to that of a conventional logistic regression model.