Comparing Low and High-Temporal Resolution DCE- MRI Texture Analysis for Discrimination of Breast Lesions from Background Enhancement

Yufeng Liu Zhejiang Chinese Medical University Jiaying Li Zhejiang Chinese Medical University Jingjing Qu Zhejiang Chinese Medical University Rui Tang Zhejiang University School of Medicine Second A liated Hospital Kun Lv Huashan Hospital Fudan University Chundan Wang Zhejiang Chinese Medical University Fengchun Xiao Zhejiang Chinese Medical University Jiali Zhou Ningbo First Hospital Min Ge Zhejiang Chinese Medical University Xuewei Ding Zhejiang Chinese Medical University Hong Ding Zhejiang Chinese Medical University Shiwei Wang Zhejiang Chinese Medical University Maosheng Xu (  xums166@zcmu.edu.cn ) The First A liated Hospital of Zhejiang Chinese Medical University; The First Clinical Medical College of Zhejiang Chinese Medical University https://orcid.org/0000-0002-2396-1600


Page 4/18
Background Breast cancer is the second leading cause of cancer death among women and is the most common malignant tumors in women worldwide [1]. In China, its incidence and mortality are increasing while the age of onset is decreasing [2,3]. Around 1.6 million people are diagnosed each year and 1.2 million succumb to their illness. Therefore, early and accurate diagnosis is essential to guide the treatment of breast cancer.
Traditional contrast-enhanced magnetic resonance imaging (CE-MRI) plays a central role in the diagnosis of breast pathologies due to its non-invasive and radiation-free nature. Yet, background enhancement (BE) may obscure enhanced lesions or contribute to false-positive readings [4][5][6]. Routine breast CE-MRI has high sensitivity and low speci city, which limits its diagnostic utility. In view of the heterogeneity of benign and malignant diseases and BE tissues, high-resolution dynamic contrast-enhanced MRI (DCE-MRI) can better evaluate the tissue microenvironment and texture characteristics. DCE-MRI can predict the biological characteristics of breast cancer, tumor environment, situation [16][17][18], molecular receptor subtypes [10,[17][18][19][20], and even genomics [9,21,22]. Quantitative assessment of ROI signal strength, spatial structure and other image parameter characteristics by high-throughput extraction can also offer diagnostic value [7,8]. For example, the quantitative analysis of the gray level cooccurrence matrix of heterogeneity may be a sign of metastasis and contribute to poor prognosis [9][10][11][12][13][14][15].
There is evidence to suggest that factors such as texture evaluation of vasculature, lymphatic permeation, density and angiogenesis are correlated with the prognosis of the disease [26]. Therefore, high-resolution DCE-MRI analysis of texture features can help predict disease prognosis [19,[23][24][25]. Such utility can help in the early diagnosis of breast cancers [4,10,27]. Therefore, in this study, we aimed to investigate the application value of texture analysis based on kinetic parametric maps from breast DCE-MRI for the discrimination of benign lesions, malignant lesions and background parenchymal enhancement using a two-compartment Extended Tofts model and volume of interest.  The contrast agent was Gadopentetate dimeglumine (Omniscan, GE Healthcare) at the dosage of 0.1 mmol/kg administered by intravenous bolus injection using a double cylinder high pressure injector (MR injection system) at a ow rate of 2.0 ml/sec,followed by a 10 ml saline ush.

Image analysis
All DCE-MRI image data was transferred into the Omni-Kinetic software (GE Healthcare, version 2.10) to obtain the K trans texture features and the pharmacokinetic parameters (K trans, K ep , V e and V p ) of the lesion central area (Lesion), surrounding peripheral area (Peri) and background enhancement (BE) area in the malignant and the benign groups using a two-compartment extended Tofts model and three-dimensional volume of interest. All data were analyzed by 2 senior imaging diagnostic radiologists (J.L. with 5 years of experience, and J.Z. with 7 years of experience in breast MRI).
The ROI (region of interest) was manually selected and delineated based on enhanced T1-weighted images and the pharmacokinetic maps. In Fig. 2, a malignant lesion was clearly shown, and a polygon ROI was drawn manually according to the shapes of lesion (lesion area), the surrounding peripheral area (peri, with a radius of 2.5-5.0 mm depending on pixel size) and background enhancement (BE) area (normal breast background parenchyma enhancement) avoiding the necrotic area and vascular beds.
Hemodynamic parameters included K trans (ml/min, endothelial transfer constant refers to the rate of blood leakage to the extravascular extracellular uid gap (EES)), K ep (ml/min, re ux rate constant refers to the blood leakage rate from EES back to blood vessels), V e = K trans /K ep (ml/ml, fractional EES volume refers to the method that fractional EES volume in the EES of the contrast agents), and V p (ml/ml, fractional plasma volume refers to the plasma volume of contrast agent) was used to quantitatively evaluate the microcirculation characteristics of the lesions.
Texture feature analysis refers to image conversion and quantitative analysis. The image conversion functions to decompose the conventional image into basic components such as space and frequency. Quantitative analysis included the structure, model, statistics and spectrum of various basic components.
For each patient, all texture features were obtained based on the K trans map from DCE-MRI images.
Texture feature analysis based on high resolution DCE-MRI included the rst-order distribution statistics, gray-level co-occurrence matrix (GLCM) and gray-level run-length matrix (GLRLM).
The rst-order distribution statistics included energy, skew, maximum, median, mean, mean absolute deviation, range, root mean square, standard deviation, variance, etc. The gray-level histogram is a discrete function of image gray level (formula: H(i) = ni/N(i = 0, 1,..., L -1), where i represents the grayscale level, L represents the number of grayscale categories, ni represents the number of pixels with grayscale level i in the image, and N represents the total number of pixels in the image), which refers to the probability of occurrence of pixels in the image with grayscale i. GLCM is a statistical method based on the second-order combination condition probability density function to describe texture, local mode, random, and spatial statistical characteristics, and to represent regional consistency and inter-regional relativity. GLRLM is the joint probability density of two position pixels, which re ects the brightness distribution characteristics and the second-order statistical characteristics of the location distribution characteristics between pixels with the same or similar brightness.

Statistical analysis
All statistical analysis was carried out in IBM SPSS Statistics software (version 19.0; BM SPSS, Chicago, 1L, USA), with P < 0.05 statistical signi cance. The Student's t-test was used to compare the differences of hemodynamic parameters from high-resolution and low-resolution DCE-MRI of the lesion central area, surrounding peripheral area and background enhancement in the malignant and benign groups (multiple correction by Family-wise error rate). Logistic regression was used to build models based on texture features for the diagnosis of breast lesions and background enhancing e ciency. ROC analysis was used to determine the diagnostic performance of each parameter.
Since the sample size in the benign and malignant groups were not uniform, over-sampling in the smaller group was conducted using the Smote (Synthetic Minority Oversampling Technique) method. For each missing sample in the smaller group, the Euclidean distance was used as the standard to calculate the distance to all samples in a small sample set, and the nearest neighbor was designated as K. A sampling ratio was set according to the sample imbalance ratio to determine the sampling ratio N. For each minority sample X, a number of samples were randomly selected from its K neighbor, assuming that the selected neighbor was Xn. For each randomly selected nearest neighbor Xn, the new sample was constructed with the original sample respectively according to the following formula: A new sample of high-resolution DCE-MRI group was constructed with the following: 62 cases --->96 cases; a new sample of low resolution DCE-MRI group was constructed with the following: 78 cases --->128 cases.
The samples were randomly divided into the training and testing groups by 0.7:0.3. As the potential correlations among many features easily affect the accuracy of the model, we conducted a three-step dimensionality reduction for the obtained features. In the rst step, the single-factor method was used to analyze and select the features with differences; in the second step, the features related to dependent variables were selected through a general linear model; in the third step, feature dimension reduction was used by the LASSO (Least absolute shrinkage and selection operator) method. After dimension reduction, the model was built by logistic regression, and the classi cation accuracy of the test group was veri ed. The area under the roc curve (AUC) was calculated. The Delong-test was used to compare the ROC curves of the models.

Study population
There were a total of 62 patients who underwent preoperative high temporal resolution DCE-MRI (1 + 26 phases) scans. 39 malignant lesions (age range, 31-77 years; average age, 54.1 years; 17 cases located in the right breast, 22 cases located in the left) and 23 benign lesions (age range, 25-79 years; average age, 48.0 years) were identi ed. In the control group, 78 patients underwent preoperative low temporal resolution DCE-MRI (1 + 6 phases) scans, which demonstrated 46 malignant (age range, 32-70 years; average age, 50.0 years; 21 cases located in the right breast, 25 cases located in the left) and 32 benign lesions (age range, 23-70 years; average age, 43.5 years). Table 1 showed general characteristics in this study.   In the malignant group, the correlation between hemodynamic parameters and pathological grade (Grade I to III) of invasive ductal carcinoma (IDC) was further analyzed. It was found that the lesion K trans -mean of the high-temporal resolution DCE-MRI group was signi cantly correlated to the pathological grading (r = 0.400, P = 0.012). In contrast, the low-temporal resolution DCE-MRI group had no correlation with the pathological grade (r = − 0,012, P > 0.05).

Performance of K trans texture feature models based on two temporal resolution DCE-MRI
According to the classi cation model constructed of the texture features based on K trans map from DCE-MRI, the AUC, accuracy, sensitivity and speci city of models from the high-temporal resolution DCE-MRI (1 + 26 phases) group were higher than those of the low-temporal resolution DCE-MRI (1 + 5 phases) group (Table 3, Fig. 3).

Discussion
In this study, we conducted a randomized controlled study on preoperative high-temporal resolution and low-temporal resolution DCE-MRI texture features, including the lesion central area, surrounding peripheral area, and background enhancement area. The quantitative parameters were measured by volume measurement, and the ROI area was selected to be larger, which was more comprehensive and accurate than the data measured at a single level in most previous studies. Neoadjuvant chemotherapy patients were excluded to avoid the effect of therapy on lesions. These results in this study indicated that high-temporal resolution DCE-MRI may be more helpful than low-temporal resolution DCE-MRI in the differentiation of breast disease from background enhancement.
Most previous studies have focused only on the disease itself [8-18, 21, 22] or the peripheral interstitial areas [19,23,24]. The lesion center, periphery and background enhancement area may be related to the differentiation of benign and malignant lesions. Few studies have comprehensively focused on the aforementioned area by volume measurement. The region of interest in this study included the lesion central area, surrounding peripheral area and background enhancement area. Therefore, the data were as comprehensive. It was found that the three ROI's hemodynamic characteristics (K trans , K ep , V e , V p , TTP) obtained from two different temporal resolution DCE-MRI had a statistical difference. Particularly, the joint K trans texture features of three areas may be helpful for the differentiation of benign and malignant breast disease. These results indicated that the comprehensive measurement of lesions and background enhancement area could be more accurate in the diagnosis and differentiation of benign and malignant lesions.
In this study, the ROC curve showed that models from the high-temporal resolution DCE-MRI group were slightly higher than those of the low-temporal resolution DCE-MRI group, which meant that DCE-MRI with high temporal resolution had high application value in the diagnosis of breast diseases. Conventional breast dynamic enhanced MRI mainly observed the characteristic of time signal intensity curve of breast disease (i.e curve shape, time to peak, early enhancement rate, etc.) and allowed for semiquantitative analysis of tumor characters. But these semiquantitative parameters were in uenced by the cardiac output, imaging sequences, contrast medium injection rate, blood ow and so on, which were prone to error [30]. This study used quantitative analysis based on breast DCE -MRI by the Extended Tofts model and obtained multiple hemodynamic parameters, such as the endothelial transfer constant of K trans , re ux rate of K ep , fractional EES volume of V e and fractional plasma volume of V p . The extended Tofts model was applicable to the time resolution of < 12 seconds, as high temporal resolution can more accurately measure the lesion and observe subtle pharmacokinetic changes in tissue for a very short time. This allows the sequence to accurately capture characters of the lesion, even including subtle differences of contrast agent concentration changes. Thus, high-temporal resolution may extract more texture features than low-temporal resolution for accurate diagnosis of breast lesions [31,32].
Although pathology is the gold standard for disease analysis and classi cation, it is limited by invasiveness and local sampling, which may cause inaccurate pathological results and the limited pathological classi cation of diseases. DCE-MRI has become an important method of diagnosis and observation of breast diseases due to its noninvasive, radiation-free and high-resolution nature with soft tissue. In this study, we analyzed the correlation between DCE-MRI and pathological grading of malignant lesions (invasive ductal carcinoma, IDC). The results showed that only the K trans -mean of lesion central area on the high-temporal resolution DCE-MRI was signi cantly correlated with pathological grade [33]. In this study, invasive ductal carcinoma was only included in the malignant group because the number of patients with other malignant types was too small (less than 5 cases) to be excluded by correlation statistics of pathological grade. These indicated that high-temporal resolution DCE-MRI may offer more diagnostic information in a lesion.
It was determined that K trans -mean re ects the hemodynamic characteristics of the disease and was related to the number and immaturity degree of blood vessels in the central tumor and peritumor regions in different pathological grades. Tumor angiogenesis increased in most malignant lesions, and there were a large number of immature tumors with high permeability of their vessel walls. Therefore, low molecular weight contrast agents were able to enter the EES through the thin vessel walls. Some previous studies also support that high-temporal resolution DCE-MRI can more accurately evaluate the hemodynamic microenvironment of tissues, through high-throughput extracted image texture features and quantitative evaluation. For example, the gray level co-occurrence matrix with the quantitative enhancement of heterogeneity can predict breast cancer invasion, prognosis and curative effect [7,[9][10][11][12][13][14][15].
Interestingly, this study found that the performance of K trans texture feature models of lesion, peripheral and background enhancement area based on high-temporal resolution DCE-MRI had more slightly high than low resolution, particularly joint textural feature model of three areas better than others. Background enhancement (BE) of the breast re ects an increase of T1 relaxation after enhancement, which directly re ects the blood supply and permeability of breast tissue. Hormones, especially estrogen, increase the microvascular permeability and vasodilation of breast tissue, causing vascular hyperplasia and ductal gland epithelial proliferation. Progesterone increases metabolic activity by promoting mitosis and leads to increased perfusion of the breast tissue, resulting in background enhancement of the breast. Tissue enhancement in MR imaging is related to vascular distribution, the permeability of contrast agents and T1 relaxation of the tissue [28]. In breast tissue, the anatomy of the vascular system and the effect of hormones on breast tissue are factors affecting the morphology and degree of BE [29]. High resolution dynamic enhanced MR texture can quantitatively evaluate disease and tissue heterogeneity, which is helpful for accurate assessment and early diagnosis of breast diseases and BE [4,10,27]. Thus, not only can high-temporal resolution DCE-MRI provide a high application value in the diagnosis of breast disease and background enhancement, but also allows for joint texture features of the lesion.
There were several limitations in this study: First, the sample size was relatively small. When the number of texture features to be extracted and screened was large, the sample size was challenged, which needs further study with large samples. Second, malignant breast lesions of other pathological types of the breast were not included in the group, which needs to be further studied. Third, high time resolution DCE-MRI scan still needs to be further improved to further optimize the scanning time. Fourth, the textural features model in this study was not further outside this center. Therefore, a multi-center validation can be implemented and add clinical characteristics to the model in the future.

Conclusions
In summary, quantitative analysis with high temporal resolution DCE-MRI has slightly high application value in the diagnosis of breast lesions and background enhancement. The joint texture analysis of the lesion, peripheral area and BE area based on high temporal resolution DCE-MRI may more helpful in the diagnosis, and the K trans -mean parameter may contribute to the pathological grading of malignant tumors. Thus, the use of texture analysis based on high temporal resolution DCE-MRI may potentially improve breast cancer diagnostic performance. All procedures performed in studies involving human participants were in accordance with the ethical standards of the institutional and/or national research committee and with the 1964 Helsinki declaration and its later amendments or comparable ethical standards.
The Ethics Committee of the First A liated Hospital of Zhejiang Chinese Medical University has approved this study, in which informed consent was written, and patient con dentiality was protected.

Consent for publication
Not applicable Availability of data and materials: The data that support the ndings of this study are available from the corresponding author upon reasonable request.