Texture Analysis of DCE-MRI Intratumoral Subregions to Identify Benign and Malignant Breast Tumors

Purpose To evaluate the potential of the texture features extracted from dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI) intratumoral subregions to distinguish benign from malignant breast tumors. Materials and Methods A total of 299 patients with pathologically verified breast tumors who underwent breast DCE-MRI examination were enrolled in this study, including 124 benign cases and 175 malignant cases. The whole tumor area was semi-automatically segmented on the basis of subtraction images of DCE-MRI in Matlab 2018b. According to the time to peak of the contrast agent, the whole tumor area was partitioned into three subregions: early, moderate, and late. A total of 467 texture features were extracted from the whole tumor area and the three subregions, respectively. Patients were divided into training (n = 209) and validation (n = 90) cohorts by different MRI scanners. The least absolute shrinkage and selection operator (LASSO) method was used to select the optimal feature subset in the training cohort. The Kolmogorov-Smirnov test was first performed on texture features selected by LASSO to test whether the samples followed a normal distribution. Two machine learning methods, decision tree (DT) and support vector machine (SVM), were used to establish classification models with a 10-fold cross-validation method. The performance of the classification models was evaluated with receiver operating characteristic (ROC) curves. Results In the training cohort, the areas under the ROC curve (AUCs) for the DT_Whole model and SVM_Whole model were 0.744 and 0.806, respectively. In contrast, the AUCs of the DT_Early model (P = 0.004), DT_Late model (P = 0.015), SVM_Early model (P = 0.002), and SVM_Late model (P = 0.002) were significantly higher: 0.863 (95% CI, 0.808–0.906), 0.860 (95% CI, 0.806–0.904), 0.934 (95% CI, 0.891–0.963), and 0.921 (95% CI, 0.876–0.954), respectively. The SVM_Early model and SVM_Late model achieved better performance than the DT_Early model and DT_Late model (P = 0.003, 0.034, 0.008, and 0.026, respectively). In the validation cohort, the AUCs for the DT_Whole model and SVM_Whole model were 0.670 and 0.708, respectively. In comparison, the AUCs of the DT_Early model (P = 0.006), DT_Late model (P = 0.043), SVM_Early model (P = 0.001), and SVM_Late model (P = 0.007) were significantly higher: 0.839 (95% CI, 0.747–0.908), 0.784 (95% CI, 0.601–0.798), 0.890 (95% CI, 0.806–0.946), and 0.865 (95% CI, 0.777–0.928), respectively. Conclusion The texture features from intratumoral subregions of breast DCE-MRI showed potential in identifying benign and malignant breast tumors.


INTRODUCTION
Breast cancer is one of the most common cancers and the main cause of cancer deaths in women, accounting for approximately 30% of new cancer cases in women and 14% of cancer deaths (1). Advances in medical technology have resulted in a relatively high cure rate for early breast cancer through radiotherapy, chemotherapy, and surgery (2,3). The treatment options for benign and malignant breast tumors differ, as do the local recurrence and survival rates (4). Benign breast tumors are generally curable through active treatment, whereas malignant tumors are difficult to cure and usually require surgery after neoadjuvant therapy to suppress local recurrence (5)(6)(7). Therefore, distinguishing benign from malignant breast tumors quickly and accurately is important.
Magnetic resonance imaging (MRI) is a non-invasive imaging method increasingly being used to detect and diagnose breast cancer. MRI has a higher sensitivity for the detection of breast lesions than mammography or breast ultrasound (8,9). Among the available MRI methods, dynamic contrast-enhanced MRI (DCE-MRI) can provide tumor anatomical information and hemodynamic information with high spatial resolution, and it plays an important role in the diagnosis, differential diagnosis, and treatment response assessment of breast cancer (10)(11)(12)(13). However, many benign lesions show strong contrast enhancement, which can lead to false-positive diagnoses, unnecessary biopsies, or overtreatment (14). The rate of preoperative breast DCE-MRI examinations is increasing, and an effective method for characterizing enhanced lesions is crucial to improve the accuracy of diagnosis.
Texture analysis refers to the extraction of texture feature parameters through specific image processing technology to obtain a quantitative or qualitative description of the texture (15,16). Texture analysis is applied to breast MRI through image processing methods, which can be used to quantify the heterogeneity of lesions (17,18). Studies have shown that texture features that characterize intratumoral heterogeneity can help identify benign and malignant breast tumors and distinguish molecular subtypes of breast cancer (19)(20)(21).
Previous studies have mainly extracted texture features from the whole tumor area in MRI images. However, the texture features derived from subregions within the breast tumor may provide valuable information to aid in clinical diagnosis and help patients develop personal treatment plans (22)(23)(24)(25). Fan et al. (26) have shown that the texture features extracted from intratumoral subregions of DCE-MRI can be used to predict Ki-67 status in estrogen receptor (ER)-positive breast cancer. To our knowledge, no research has been performed on the identification of benign and malignant breast tumors on the basis of texture features extracted from intratumoral subregions of breast DCE-MRI. The purpose of this study was to evaluate the potential of the texture features extracted from DCE-MRI of intratumoral subregions for distinguishing benign and malignant breast tumors.

Study Cohort
This study was approved by the Ethics Review Committee at Shengjing Hospital of China Medical University (No. 2019PS175K), and the requirement for informed consent was waived because of the retrospective nature of the study. Between January 2017 and January 2020, patients who underwent breast DCE-MRI examinations were reviewed through the image archiving and communication system (PACS) at our institution. The study cohort initially included 378 patients. The inclusion criteria were as follows: (1) patients who underwent breast DCE-MRI and (2) patients with benign or malignant breast tumors confirmed by histopathology. The exclusion criteria were as follows: (1) patients treated with surgery, chemotherapy, or radiotherapy before DCE-MRI (n = 43); (2) patients diagnosed through excisional biopsy before DCE-MRI (n = 26); and (3) patients with insufficient image quality for subsequent processing because of obvious motion artifacts (n = 10). Consequently, 299 patients (mean age, 48.30 ± 9.74 years; range, 25-84 years) were divided into training (n = 209) and validation (n = 90) cohorts by different MRI scanners, including 124 benign and 175 malignant breast tumors. The clinical characteristics of the study cohort are summarized in Table 1. The flowchart of this study is shown in Figure 1.

Image Acquisition
In the training cohort, DCE-MRI examinations were performed with a GE 3.0T MRI scanner (Signa HDxt, GE Healthcare) equipped with a dedicated eight-channel bilateral breast coil on patients in a prone position. A transverse fat-suppression T1weighted pre-contrast scan was first obtained with the VIBRANT-VX technique. Eight phases of fat-suppression T1weighted post-contrast scans were acquired after intravenous injection of the contrast agent (Magnevist, Bayer-Schering Pharmaceuticals, Germany) at a dose of 0.15 mmol per kg body weight at 4 mlL/s and subsequent flushing with an equal volume of saline at the same injection speed. The following imaging parameters were used: repetition time (TR), 7.42 ms; echo time (TE), 4.25 ms; flip angle, 15°; slice thickness, 2.20 mm; spacing between slices, 2.20 mm; field of view, 340 × 340 mm 2 ; image matrix, 1,024 × 1,024; slice number, 78. For each patient, eight phases of subtraction images were obtained by subtracting pre-contrast images from eight post-contrast images.
In the validation cohort, DCE-MRI examinations were performed with a Philips 3.0T MRI scanner (Ingenia, Philips Medical System, Best, Netherlands) equipped with a dedicated seven-channel bilateral breast coil with patient in a prone position. First, an axial fat-saturated T1-weighted pre-contrast scan was acquired. Then, eight axial contrast-enhanced fatsaturated T1-weighted scans were acquired after the intravenous bolus injection of the same contrast with the same dose. The imaging parameters were as follow: repetition time (TR), 4.14 ms; echo time (TE), 2.10 ms; flip angle, 12°; slice thickness, 2.00 mm; spacing between slices, 1.00 mm; field of view, 340 × 340 mm 2 ; matrix, 380 × 380; slice number, 78. Eight subtraction sequences were obtained by subtracting the precontrast scan from each of the eight post-contrast scans.

Image Processing and Lesion Segmentation
Two senior radiologists, with 10 and 15 years of experience in interpreting breast MRI were invited to review the subtraction images in the fourth phase and reached a consensus in selecting the slice image with the maximum tumor diameter for each patient for subsequent analysis (27). During the image review, the radiologists were blind to the patients' pathological results. The whole tumor area was segmented with a semi-automatic method in Matlab 2018b (Mathworks, Natick, MA, USA), as described below (28,29). One of the two radiologists manually delineated a region of interest (ROI) with an arbitrary shape around the lesion area on the subtraction image. The pixel gray levels within the ROI were first normalized to m ± 3s (m: mean gray level of pixels within the ROI; s: standard deviation), and the range was quantized to 8 bits/pixel to change the signal to noise ratio of the texture results (30)(31)(32). A spatial fuzzy C-means (FCM) algorithm was then used to delineate the contour boundary of the lesion according to the ROI, and the whole lesion area was refined through morphological processing methods (33)(34)(35). Another radiologist verified and proofread the results of the semi-automatic breast tumor segmentation.

Intratumoral Subregion Partition
To better understand the intratumoral heterogeneity of breast tumors, as in a previous study (26), we divided the lesion area into three subregions according to the variations in pixel signal intensity in different phases. The specific partition details are as follows: The relative enhancement of the post-contrast image compared with the pre-contrast image on a pixel-by-pixel basis was calculated with the following formula: where I(m, n, t) and I(m, n t 0 ) represent the signal intensity of the pixel (m, n) captured at times t and t 0 (the pre-contrast moment) (36). The time-signal intensity curve, H(m, n, t), was defined to describe the variation in the relative enhancement over time (37)(38)(39). The time to peak (TTP), which represents the arrival time of the peak relative enhancement, was calculated with the following formula: Then the pixels within the tumor region were divided into three subregions according to their TTP values. More specifically, pixel sets at the first four, fifth or sixth, and seventh or eighth phases to achieve peak enhancement values were defined as early, moderate, and late subregions, respectively; this method was similar to those described in previous studies  (26,36). Therefore, the tumor was divided into three regions representing different sets of TTP values.

Texture Feature Extraction
A total of 467 texture features were extracted from the whole tumor area and the three subregions with Matlab 2018b. The feature extraction methods could be classified into the following four categories: histogram, gray-level co-occurrence matrix (GLCM), gray-level run length matrix (GRLM), and discrete wavelet transform (DWT). Detailed information on the features is shown in Table 2. Each GLCM feature was calculated by using four angles (0, 45, 90, and 135°) and four distances (1, 2, 3, and 4 pixels). Each GRLM feature was calculated by using four angles (0, 45, 90, and 135°) and a distance of 1 pixel. In the following,

Feature Selection and Model Construction
To reduce the dimensionality of the features, the correlation between features was first tested with Pearson's correlation analysis, and features with correlation coefficients of >0.95 relative to other features were removed. The remaining features were filtered by the least absolute shrinkage and  selection operator (LASSO) method to select the optimal feature subset (40). Two machine learning models, decision tree (DT) and support vector machine (SVM), were used to construct classification models based on the optimal feature subset in the training cohort with a 10-fold cross-validation method for identifying benign and malignant breast tumors. And the classification models were tested by using a independent validation cohort. The 10-fold cross-validation refers to random division of the data set into 10 sets, nine of which were used for training and the last of which was used for testing. This process was repeated 10 times, and the test data differed each time.

Statistical Analysis
All statistical analyses were performed in SPSS 22.0 (IBM, Armonk, NY, USA). The Kolmogorov-Smirnov test was first performed on texture features selected by LASSO to assess whether the samples followed a normal distribution (41); if so, the variables in the tables are represented by means ± standard deviation (SD), and if not, the variables in the tables are represented by medians ± interquartile range. Univariate logistic regression analysis was used to evaluate the performance of an independent feature in distinguishing benign from malignant breast tumors. The receiver operating characteristic (ROC) curve constructed in the professional statistics software MedCalc (version 14.10.20, http://www.medcalc.org/) was used to assess the classification performance by calculating the area under the ROC curve (AUC). The corresponding accuracy, sensitivity, and specificity were also determined. The DeLong test was used to determine the statistical significance of differences between AUCs. A two-tailed P value of <0.05 was considered statistically significant. The intraobserver variability of texture features extracted by the two radiologists was evaluated by using intraclass correlation coefficients [ICC, (0, 0.4), poor agreement; (0.4, 0.6), moderate agreement; (0.6, 0.8), good agreement; and (0.8, 1), excellent agreement] (42,43).

Study Cohort
A total of 299 patients were enrolled in this study. In the training cohort, the patients had 84 (40.2%) benign breast tumors classified into three histopathological types: adenosis (48), fibroadenoma (32), and papilloma (4). The 125 (59.8%) malignant breast tumors comprised 116 invasive carcinomas of no special type, 6 ductal carcinomas in situ, 2 invasive micropapillary carcinomas, and 1 invasive lobular carcinoma. In the validation cohort, the patients had 40 (44.4%) benign breast tumors classified into three histopathological types: adenosis (23), fibroadenoma (14), and papilloma (3). The 50 (55.6%) malignant breast tumors comprised 41 invasive carcinomas of no special type, 5 ductal carcinomas in situ, 3 invasive micropapillary carcinomas, and 1 invasive lobular carcinoma. The results of the whole tumor area segmentation and intratumoral subregion partition are displayed in Figure 2, which shows two randomly selected cases, one benign case and the other malignant case.

Univariate Analysis
The results of univariate logistic regression analysis for identifying benign and malignant breast tumors are displayed in Table 3, which shows the top six features with the best performance extracted from the three subregions and the whole tumor area. The AUCs of features derived from the whole tumor area ranged from 0.732 to 0.786. Features from the early subregion performed best among the three subregions, with AUC values ranging from 0.787 to 0.886. The AUCs of the run length non-uniformity (1, 0) (P < 0.001), difference square (0, 1) (P = 0.004), and short run emphasis (1, 0) (P < 0.001) from the early subregion were significantly higher than those from the whole tumor area. The AUCs from the moderate subregion ranged from 0.715 to 0.777, and the AUCs from the late subregion ranged from 0.685 to 0.884. Among all individual features, the run length nonuniformity (1, 0) extracted from the early region achieved the highest AUC of 0.886 [95% confidence interval (CI), 0.836-0.926]. Table 4 shows the performance of the classification models for distinguishing benign from malignant breast tumors in the training and validation cohorts, and the corresponding ROC curves are presented in Figures 3 and 4. In the training cohort, the AUCs of the DT_Whole model and SVM_Whole model were 0.744 and 0.806, respectively. In contrast, the AUCs of the DT_Early model (P = 0.004), DT_Late model (P = 0.015), SVM_Early model    Table 6. However, there was no significant difference between the SVM_Early model and the SVM_Late model in the training and validation cohorts (P = 0.524 and P = 0.523, respectively), and no significant difference between the DT_Early model and the DT_Late model (P = 0.945 and P = 0.332, respectively). Fifteen texture features extracted from the early subregion and 17 features extracted from the late subregion were selected by LASSO, as listed in Table 7.

DISCUSSION
This study investigated the relationship between texture features extracted from intratumoral subregions of breast DCE-MRI and the differential diagnosis of benign and malignant breast tumors. Features from subregions were able to distinguish benign from  Texture analysis can characterize intratumoral heterogeneity on the basis of quantitative image features extracted from conventional medical imaging to help diagnose, stage, and predict the prognosis and response to treatment in multiple oncology fields (44)(45)(46). Intratumoral heterogeneity reflects differences in biological characteristics, such as gene expression, metabolism, and angiogenesis (23,47). Texture features derived from intratumoral subregions that reflect the heterogeneity of breast tumors, rather than the whole tumor area, may play a more important role in the prognostic analysis and identification of hormone receptor status in breast cancer (26,36). A previous study has shown that texture features extracted from subregions with rapid delayed washout can be used to assess ER status and lymph node classification in breast cancer (48). Chang et al. (49) have quantified intratumoral heterogeneity on breast DCE-MRI by using a subregion-based feature extraction method for predicting ER status, human epidermal growth factor receptor 2 (HER2) status, and triple-negative breast cancer, achieving accuracy of 73.53, 82.35, and 77.45%, respectively. In this study, an intratumoral subregion partition method was used to distinguish benign from malignant breast tumors. Texture   features were derived from three subregions and the whole tumor area, and the corresponding classification models were established. The models built with features from the early and late subregions achieved better performance than models built with features from the whole tumor area. A possible explanation for this finding is that the intratumoral subregions reflect angiogenesis, which may be indicative of the aggressiveness of malignant breast tumors (50).
A previous study has investigated the diagnostic performance of mammography texture analysis in differentiating benign from malignant breast tumors (51). In the present study, the subtraction images of DCE-MRI were used for texture analysis. Previous studies have discussed the roles of histograms, GLCM, and GRLM-based texture features in the differential diagnosis or treatment response assessment in breast cancer (14,44). In addition to the features used in these studies, DWT-based features were extracted in this study. DWT is used to modify the image from the spatial domain to the frequency domain and has been extensively applied to feature extraction from electroencephalogram signals (52,53). In the present univariate analysis, the DWT-based features derived from the late subregion, including Deubechies2_2VH (P = 0.001), Haar_1HH (P = 0.001), and Haar_4HH (P < 0.001), performed better in distinguishing benign from malignant breast tumors than those derived from the whole tumor area.
Two prevalent machine learning methods, DT and SVM, were applied to establish classification models in this study. To prevent overfitting, a 10-fold cross-validation method was used. The models established with features from the early and late subregions achieved better performance than models from the whole tumor area in the training and validation cohorts. However, no significant differences were found between the performance of models from the moderate subregion and that of models from the whole tumor area in the training and validation cohorts (P = 0.473 and P = 0.078, P = 0.406 and P = 0.664, respectively). Furthermore, the SVM_Early model, SVM_Moderate model, and SVM_Late model had higher AUCs than the DT_Early model, DT_Moderate model, and DT_Late model. SVM initially maps the input vector to a higherdimensional feature space and identifies the hyperplane that divides the data points into two categories; the resulting classifier can reliably classify new samples and achieve considerable versatility (54).
A previous study by Li et al. (55) has applied four methods to classify benign and malignant breast tumors, and reported that the DT model achieved the best performance, with an AUC of 0.781, a sensitivity of 0.6, and a specificity of 0.894. Another study has used The data are medians ± interquartile range.
an SVM model for classifying benign and malignant breast tumors and obtained a sensitivity of 66.67% and a specificity of 93.55% (56). Wang et al. (20) have used logistic regression analysis to distinguish benign and malignant breast tumors, and achieved an accuracy of 79.5%, a sensitivity of 0.607, a specificity of 0.800, and an AUC of 0.802. In comparison, the best classification performance of our SVM_Early model achieved an AUC of 0.934, a sensitivity of 89.6%, a specificity of 86.9%, and an accuracy 88.5%. However, studies in which the classification model is based on deep learning methods have reported higher accuracy in distinguishing benign and malignant breast lesions (57,58). In addition, we separately evaluated the intraobserver variability of texture features extracted from the whole lesion region and from three different intratumoral subregions. The two radiologists showed high consistency in calculating texture features from the single-slice method, and all ICCs were greater than 0.8. The intraobserver variability was mainly related to slice selection and ROI delineation. Hence, standardized strategies for ROI determination are crucial.
This study has some limitations. First, the sample size was relatively small. Second, only a representative single-slice image was analyzed, and thus some useful information on the tumor might have been missed. Texture analysis based on threedimensional breast tumor lesions may yield more useful information (59). Finally, the subtraction images of breast DCE-MRI were used to extract texture features. Features derived from post-contrast images or diffusion weighted imaging images may be helpful in distinguishing benign from malignant breast tumors (60).

CONCLUSION
The texture features extracted from intratumoral subregions of breast DCE-MRI can be used as imaging biomarkers for the differential diagnosis of benign from malignant breast tumors. Specifically, features derived from subregions representing the early and late TTP values achieved better performance than features from the whole tumor area. Further research with a larger sample size is needed to verify the results of this study.

DATA AVAILABILITY STATEMENT
The original contributions presented in the study are included in the article/supplementary material. Further inquiries can be directed to the corresponding author.

ETHICS STATEMENT
The studies involving human participants were reviewed and approved by Shengjing Hospital of China Medical University. Written informed consent for participation was not required for this study in accordance with the national legislation and the institutional requirements. Written informed consent was not obtained from the individual(s) for the publication of any potentially identifiable images or data included in this article.