The present study was approved by the Institutional Review Board of Zhengzhou University and waived the requirement for informed consent due to its retrospective nature. 522 patients with HCC, hepatic metastases, hepatic hemangioma, hepatic cysts, FNH, HA, underwent CT examination at the First Affiliated Hospital of Zhengzhou University from 2016 to 2018, were included.

The following inclusion criteria were used: (1) HCC, hepatic metastases, FNH, and HA are histopathologically confirmed by surgery or liver percutaneous needle biopsy; (2) enhanced CT scan was performed 30 days before surgery or biopsy; (3) liver occupying lesions number was <= 3. The exclusion criteria were as follows: (1) insufficient quality of CT imaging; (2) lesions less than 1.0 cm (to avoid the impact of small volume effect of the lesion); (3) patients with multisource cancer; (4) patients with hepatitis B, cirrhosis, liver injury or liver surgery found on non-contrast CTs. Seventy patients were excluded, which finally resulted in the enrollment of 452 patients: 77 with HCC, 104 with liver metastases, 126 with hepatic hemangioma, 99 with hepatic cysts, 24 with FNH, 22 with HA. All these lesions were included in the test group.

Hepatic hemangioma, hepatic cysts were confirmed by referencing radiologic reports by experienced radiologists and adhering to the following criteria (14): hepatic hemangioma (enhanced nodules around the arterial phase, increased centripetal filling during the venous phase, uniform enhancement during the arterial phase, or continuous enhancement during venous phase); hepatic cyst (close to the CT value of water and no apparent contrast enhancement).

Control group: 452 lesions were diagnosed on unenhanced CT images by an attending physician who worked for 5 years and a chief physician who worked for more than 10 years without knowing the histopathological examination results. The diagnostic accuracy was evaluated with histopathological examination results (HCC, hepatic metastases, FNH, HA) and radiologic reports (hepatic hemangioma, hepatic cyst) as the gold standard.

In addition, six kinds of focal hepatic lesions were divided into benign and malignant groups. The benign group included hepatic hemangioma, hepatic cyst, FNH and HA; the malignant group included HCC and hepatic metastases.

All examinations were acquired on spectral CT scanners (first generation of Discovery CT750 HD scanner; the second generation of the Discovery CT scanner; both from GE Medical Systems). The following specific scanning parameters were adopted: tube voltage of 120 kV, tube current of 240–300 mA, and layer thickness of 5 mm. Enhanced scanning was performed by injecting a non-ionic contrast agent (iodine content 320 g/L) into the cubital vein at 2.5–4.0 mL/s; the total dose calculated was 1.5 mL/kg. After the contrast agent was injected, the arterial phase was scanned 25–30 s, the venous phase was scanned in 60s, and the delayed phase was scanned after 180 s.

Original Digital Imaging and Communications in Medicine (DICOM) images were imported to the Radcloud platform (Huiying Medical Technology Co., Ltd., Beijing, China) for radiomic features extraction and analysis (15). One focal hepatic lesion (not adjacent to the inferior vena cava, caudate lobe, the great hepatic vein, artery, and bile duct) was selected for each patient. Two radiologists manually delineated the region of interest (ROI) along the edge of the lesion, layer by layer, on unenhanced CT images. The volume of interest (VOI) of the lesion was automatically generated by the computer. Another senior radiologist examined the outline results. During the segmentation, corresponding enhanced CT images were referred to determine tumor boundary. Interclass correlation coefficients (ICCs) were used to assessed the interobserver agreement of feature extraction in order to validate the stability of the radiomics features (16). An ICC of greater than 0.75 was considered acceptable. The procedure of radiomics method is illustrated in Figure 1.

A total number of 1409 quantitative imaging features were extracted from CT images on the Radcloud platform. These features were grouped into four groups. Group 1 (first-order statistics) consisted of 18 descriptors that quantitatively delineated the distribution of the voxel intensities for CT images through commonly used basic metrics. Group 2 (shape- and size-based features) contained 14 three-dimensional features that reflect the shape and size of the region. Calculated from grey-level run-length and grey level co-occurrence texture matrices, 75 textural features that could quantify the region heterogeneity differences were classified into group 3 (texture features). Finally, group 4 (higher-order statistics features) included the first-order statistics, and texture features were derived from wavelet transformation of the original image. We used seven types of filters that included a total number of 14 filters: exponential, gradient, square, square root, logarithm, lbp-2D, and wavelet (wavelet-LLL, wavelet-HHH, wavelet-HLL, wavelet-HHL, wavelet-LLH, wavelet-HLH, wavelet-LHL, and wavelet-LHH).

To explore the generalization performance of the model, training and testing set were split based on a ratio of 8:2 using computer-generated random numbers. All radiomic features were standardized using the Standard Scaler function by removing the mean value, followed by division by its standard deviation and transformation into feature values with zero as the mean and one as the standard deviation.

LASSO is a regression analysis method that can perform both variable selection and regularization to improve the identification accuracy and interpretability of the model. For example, it has a tuning parameter to control the penalty of the linear model, which guarantees the minimum penalty when obtaining a model with a smaller number of features, where the penalty is mean square error (MSE). In addition, another parameter controls the correlation of features, making the selected features less relevant. L1 regularization was used as the cost function, the error value of cross validation was 5, and the maximum number of iterations was 1000. The optimization goal of LASSO is:

Where X is the radioactivity characteristic matrix, y is the sample vector marker, n is the sample number, w is the coefficient vector regression model, alpha∗║w║ is the lasso punishment.

In this study, we used Support Vector Machine (SVM) classifiers on the Radcloud platform to establish radiomic-based models based on the extracted optimal features. The basic principle of SVM, is to find an optimal hyperplane that produces a better generalization of the dataset. It develops a model that predicts whether a new sample falls into one of the categories or not. Let’s given a training data set S = {(x1, y1),…,(xn, yn)} where xi ∈ Rn and y i ϵ{+1,−1}.

The xi represents the transferred input vector and yi is the target value. SVM is a binary classifier in which the class labels contain only two values + 1 or −1. From the inputs, SVM draws an optimal hyper-plane H that separates the data into different classes and the hyper-plane H can be defined as:

The algorithm is based on finding the hyper-plane which gives the maximum distance of separation between training samples using the following function.

For the problem of multiclass learning, SVM solved it as a single multi-class problem further it is modified into multiple binary problems. Hence, the optimal hyper-plane can be combined by the inequality.

So, the optimization problem can be written.

Notice that the output value of the SVM is either -1 or 1. When the output value of a subimage of suspicious tumor region is 1, the system will classify the tumor in the CT image as a certain category. Conversely, when the output value is -1, the hepatic tumor will be diagnosed as other.

All statistical analysis were performed with Python 2.7 for Linux. Support vector machine (SVM) were employed based on the optimal discriminative features. Further, the model classification performance was evaluated by AUC (Area Under Curve), accuracy, overall accuracy ((true positives+ true negatives)/(true positives+ false positives+ false negatives+ true negatives)), precision (true positives/(true positives+ false positives)), recall (true positives/(true positives+ false negatives)), f1-score (precision *recall* 2/(precision + recall)), and support (total number in a test set). The formula of confidence interval was as follows:

If

If

A retrospective analysis of 452 patients (299 men and 153 women; mean age, 53 ± 14.58 years; range, 3-89 years) was included. The relevant lesions studied included 77 with HCC, 104 with hepatic metastases (40 were lung cancer with hepatic metastasis, 34 were gastric cancer with hepatic metastasis, 30 were colorectal cancer with hepatic metastasis), 126 with hepatic hemangioma, 99 with hepatic cysts, 24 with FNH, 22 with HA.

Control group: The diagnostic accuracy based on unenhanced CT images of the two radiologists was 0.51(231/452) and 0.69 (311/452) respectively.

Our feature stability analysis showed that the ICC for most of radiomics feature was high (Supplementary Tables 4, 5). Therefore, all outcomes were based on the measurements of the first radiologist. In the present study, 1409 texture parameters were obtained from each patient. The LASSO algorithm (Figure 2) was used to evaluate the correlations between each pair of texture parameters and reduce the dimensionality of the above high-dimensional features based on the optimal λ parameters, and the optimal discriminative features were screened. The final optimal discriminative feature set was comprised of 6 first-order statistics, 1 shape, 8 texture, and 11 higher-order statistics features, a total number of 27 (Table 1).

The diagnostic efficacies of SVM for the training set and testing set are presented in Table 2. Generally, SVM attained a satisfactory classification performance with an AUC range within 0.942∼1 in the training set and 0.897∼0.995 in the testing set (Figure 3). The overall accuracy in the training and testing set were 0.88 and 0.76 (Table 2, Figure 4), respectively. The other four indicators (precision, recall, f1-score, and support) are summarized in Supplementary Table 1 for the training set and in Supplementary Table 2 for the testing set.

All included patients were divided into benign group and malignant group. The benign group included hepatic hemangioma, hepatic cyst, FNH and HA (272 patients in total); the malignant group included HCC and hepatic metastases (180 patients in total). 34 optimal discriminative features were selected to distinguish the benign and malignant tumors(Supplementary Table 3, Supplementary Figure 1). Generally, SVM attained a satisfactory classification performance, and the AUC in the training and testing set was 0.951 and 0.899 (Table 3, Figure 5). The overall accuracy in the training and testing set was 0.89 and 0.84 (Table 3, Figure 6), respectively.