Comparing Texture Analysis of Apparent Diffusion Coefficient MRI in Hepatocellular Adenoma and Hepatocellular Carcinoma

Aim: This study aimed to assess the effectiveness of using MRI-apparent diffusion coefficient (ADC) map-driven radiomics to differentiate between hepatocellular adenoma (HCA) and hepatocellular carcinoma (HCC) features. Materials and methods: The study involved 55 patients with liver tumors (20 with HCA and 35 with HCC), featuring 106 lesions equally distributed between hepatic carcinoma and hepatic adenoma who underwent texture analysis on ADC map MR images. The analysis identified several imaging features that significantly differed between the HCA and HCC groups. Four classification models were compared for distinguishing HCA from HCC including linear support vector machine (linear-SVM), radial basis function SVM (RBF-SVM), random forest (RF), and k-nearest neighbor (KNN). Results: The k-nearest neighbor (KNN) classifier displayed the top accuracy (0.89) and specificity (0.90). Linear-SVM and KNN classifiers showcased the leading sensitivity (0.88) for both, with the KNN classifier achieving the highest precision (0.9). In comparison, the conventional interpretation had lower sensitivity (70.1%) and specificity (77.9%). Conclusion: The study found that utilizing ADC maps for texture analysis in MR images is a viable method to differentiate HCA from HCC, yielding promising results in identified texture features.


Introduction
Hepatocellular adenoma (HCA) is a rare benign liver tumor primarily associated with steroid use, particularly oral contraceptives, and it is more commonly found in young women [1].However, there has been a recent increase in HCA cases among males and those without steroid exposure, possibly due to the inflammatory subtype of HCA related to obesity and alcohol use [2].Proper management is crucial for HCA lesions due to the potential complications, including hemorrhage and transformation into hepatocellular carcinoma (HCC) [3].HCC is the primary malignant liver tumor that originates from hepatocytes.Early detection of HCC is crucial for effective treatment; however, the majority of cases are diagnosed at advanced stages, resulting in poor prognosis and low five-year survival rates of 2.5% [4].
While the majority of HCC cases occur in individuals with liver cirrhosis, around 20% of cases develop in non-cirrhotic patients and are typically diagnosed at advanced stages [5].Guidelines suggest utilizing fatsensitive imaging techniques (e.g., contrast-enhanced ultrasonography, CT, MRI) with contrast agents to detect HCA in liver lesions.These methods aid in accurate diagnosis and management by identifying specific components.They are also used to assess HCC and provide additional insights [6,7].However, distinguishing between HCA and HCC presents a difficulty for radiologists and clinicians due to shared histopathological and imaging features.Consequently, the use of advanced imaging techniques and radiomic analysis becomes crucial to enhance accuracy in distinguishing between HCA and HCC [8,9].Accurate diagnosis is crucial for managing HCA and HCC.Radiomics, a diagnostic tool, extracts high-throughput features from imaging across modalities, enhancing diagnostic accuracy and enabling personalized treatment strategies for HCA and HCC [6,[10][11][12].
Recent studies have prioritized analyzing radiomic features to distinguish liver lesions, especially HCC with its high prevalence and poor prognosis [13,14].Limited research has been conducted on using radiomic feature analysis to differentiate HCA and HCC using CT and MRI imaging features [15,16].However, further investigation is required to establish the effectiveness of this method and gather data for future assessments in this emerging field.This study aimed to use apparent diffusion coefficient (ADC) MRI-driven radiomics features to differentiate between HCA and HCC.The findings expand our understanding of using radiomics to distinguish between HCA and HCC liver lesions, offering a less invasive and operator-independent approach.

Materials And Methods
This prospective cross-sectional study was conducted at Imam Khomeini Hospital Complex, Tehran University of Medical Sciences, Iran, from 2020 to 2022.This study aimed to recruit patients diagnosed with HCA or HCC.Patients with clinical history and suspicious imaging features indicative of HCA or HCC were included.The final diagnosis was confirmed through histology.MRI examinations were performed for liver evaluation in enrolled patients, while those with contraindications were excluded.The study included 55 patients (27 males, 28 females).Among the 106 confirmed hepatic tumors, an equal distribution of 53 tumors each was observed for both HCC and HCA, utilizing a 3T GE healthcare whole-body system for imaging.

ROI segmentation
An experienced radiologist with 16 years of experience in liver radiology performs region of interest (ROI) segmentation on the ADC map.Care is taken to include the entire tumor volume while avoiding adjacent normal liver tissue or artifacts.

Texture analysis
The texture analysis describes the process of extracting radiomic features from segmented regions of interest (ROIs) using a custom-developed algorithm.The gray-level values of images were normalized by subtracting the mean value and dividing by the standard deviation.As suggested by radiomics experts, the gray-level values were additionally scaled by a factor of 100 and a shift of 300 to avoid negative values.PyRadiomics, an open-source Python library, was utilized to extract a total of 102 radiomic features, including first-order intensity statistics features and texture features.These features capture different aspects of tumor heterogeneity, such as intensity fluctuation, spatial patterns, and voxel interactions.

Feature selection algorithms and classification
Three feature selection algorithms and four classifiers from the "scikit-learn" Python library were combined (all possible combinations) to differentiate between benign HCA and malignant HCC tumors.The feature selection algorithms employed were as follows: recursive feature elimination (RFE) algorithm [17], sequential feature selector [18], and feature selection based on the k highest scores [19].
In this study, different classifiers were employed for distinguishing between benign HCA and HCC tumors, including support vector classifier with "linear" or "radial basis function (RBF)" kernels [20].Random forest classifier, 10 to 200 number of trees were tested to find the best number of trees [21].K-nearest neighbors classifier (KNN), various numbers of neighbors, ranging from three to 15, were evaluated to identify the optimal value [22].Grid search was utilized to find the best model and parameters, while four-fold crossvalidation was performed to assess the models' performance.ROC curves were available for every classifier.
The texture features obtained from the analysis were subjected to statistical analysis to assess their effectiveness in distinguishing between HCA and HCC.Descriptive statistics, including mean, median, standard deviation, and range, were calculated for each texture feature.Univariate analysis, employing either the Wilcoxon p-value test or Student's t-test, was conducted to compare the texture features between the HCA and HCC groups.Receiver operating characteristic (ROC) curve analysis was performed to evaluate the diagnostic performance of the selected texture features.The sensitivity, specificity, and area under the curve (AUC) were determined from the ROC analysis.

Results
A total of 102 features were extracted from the MRI assessments performed in the study.The comparison of these features between the two study groups showed statistical differences in 22 features based on the t-test and 45 features based on the Mann-Whitney U test.HCA was present in 48 images (102 features).Also, HCC was present in 51 images (102 features).
Out of the 102 extracted features, a total of 10 features for linear-SVM, 30 features for RBF-SVM, five features for RF, and 10 features for KNN.Among the four classifiers employed to analyze the radiomic data, the KNN exhibited the highest accuracy (0.89) and specificity (0.90).Also, the linear-SVM and KNN demonstrated the highest sensitivity of 0.88 for both, while the KNN classifier achieved the highest precision of 0.9 (Table 1).

Variable
Linear  The AUC serves as a measure of the model's classification performance.A higher AUC value, closer to one, indicates superior performance in distinguishing between positive and negative classes (Table 2).The diagnostic performance of MRI in distinguishing between HCA and HCC was assessed using the ROC curve.The ROC curve provided valuable information, with an area under the curve (AUC) of 0.860 (95% confidence interval: 0.783-0.936,p<0.001).In terms of discrimination between HCA and HCC, the sensitivity and specificity were determined to be 70.1% and 77.9%, respectively, as depicted in Table 4.The discrimination between HCA and HCC using the three-dimensional axial LAVA dynamic (3D AX LAVA DYN) test revealed AUC, 95% CI, p-value, sensitivity, and specificity (Table 5).

Discussion
Oncology has recently focused on using texture analysis (TA) in various imaging modalities like X-ray, ultrasound, CT, MRI, and PET.TA has been applied to investigate the connection between textural parameters and tumor pathology data [23].Holli et al. used textural metrics from breast MRI to differentiate between invasive lobular carcinoma and invasive ductal carcinoma [24].Ba-Ssalamah et al. utilized contrastenhanced CT texture parameters to distinguish between different types of gastric tumors [25].Georgiadis et al. demonstrated the importance of texture analysis (TA) in brain MRI for distinguishing between metastases, gliomas, and meningiomas [26].These studies highlight the strong correlation between textural features and pathogenic information.
The study concluded that texture analysis based on ADC MR imaging is a reliable method for distinguishing between HCA and HCC in the liver.The most significant results were observed when analyzing ADC map images, which showed distinct TA features between HCA and HCC tumors.These images achieved high sensitivity, specificity, precision, and diagnostic accuracy in accurately classifying HCA and HCC.The findings are consistent with another study by Stocker et al. that demonstrated the effectiveness of dimension texture analysis (2DTA) in differentiating malignant and benign hepatocellular tumors in the non-cirrhotic liver [27].The study also reported the sensitivity and specificity when using TA features and traditional radiological interpretation.
Malignant tumors generally exhibit higher cellularity, leading to lower ADC values in diffusion-weighted imaging (DWI) compared to benign tumors.ADC values can be influenced by various factors, such as vendor, hardware, sequence, and approach [28].Lee et al. demonstrated that DWI with an ADC map is a quantitative imaging method that is minimally affected by variations in gain factors, and several parameters derived from ADC showed significant differences between benign and malignant soft-tissue tumors [29].Another study found significant differences in first-order-based ADC parameters between intermediate and high-grade sarcoma patients using 1.5 T MRI [30].Additionally, a previous study on myxoid soft-tissue tumors found a significant difference in kurtosis of ADC between benign and malignant cases [31].While the ADC is an important quantitative parameter in diffusion-weighted imaging (DWI), qualitative analysis of DWI with low and high b values is also valuable.DWI with a high b value does not simply serve as an inverted image of the ADC but also provides distinct information.It's worth noting that not all regions with low ADC values indicate high cellularity; they could instead be attributed to fatty components or T2 blackouts caused by hematoma [32].
Our study utilized ADC map images to identify multiple TA features, including 102 features, and found statistically significant differences between HCA and HCC.The classification of these features using linear-SVM, RBF-SVM, random forest, and k-nearest neighbors (KNN) revealed that the KNN model of TA features performed better in distinguishing HCA from HCC lesions.Based on these results, the k-nearest neighbors (KNN) classifier exhibited the highest accuracy (0.89) and specificity (0.90).The reasons behind KNN's strong performance can be attributed to its ability to capture local patterns, its non-parametric nature that allows it to handle complex decision boundaries, and its reliance on nearest neighbors for classification.However, it's important to note that KNN also has some limitations, such as sensitivity to the choice of the number of neighbors (k) and computational inefficiency on large datasets, which should be considered when selecting an appropriate classifier for a specific task.In a previous study, a logistic regression model utilizing TA features from arterial-phase images achieved an accuracy of 84.5% (sensitivity 84.1%, specificity 84.9%) for HCC diagnosis.The logistic regression model demonstrated higher accuracy compared to other models reported in the literature [33][34][35][36][37].
Additionally, results from a study by Książek et al. indicated that their proposed model achieved the highest accuracy in detecting HCC [38].The study observed significant differences between HCA and HCC in certain MRI sequences, while some sequences showed no significant differences [39].Dynamic MRI proved effective in distinguishing between the two conditions, with an AUC of 0.860, indicating good discrimination.

TABLE 1 : Results of the classification of HCA and HCC tumors.
The first row shows the feature selection algorithm for each classifier and the number of selected features in parentheses.SVM: support vector machine; RBF: radial basis function; KNN: k-nearest neighbors

TABLE 2 : Area under the curve for SVM (linear), SVM (RBF), RF, and KNN.
Comparisons of mean and SD of 10 features classifiers extracted from apparent diffusion coefficient (ADC) images between HCA and HCC groups are shown in Table3.These findings suggest that the analyzed features exhibit variations between HCA and HCC, which may have diagnostic or prognostic implications in distinguishing between these two tumor types.