This article is mentioned in:

Introduction

Breast cancer is a common malignant disease in women. Besides mammography and echography, MR imaging is of great significant in breast cancer diagnosis since its higher sensitivity and specificity (1). Depends on the abundant imaging sequences, MR imaging provides a wealth of morphological information for the diagnosis of breast disease. Multiple MR sequences describe the breast tumors in different aspects, dynamic contrast enhanced magnetic resonance imaging (DCE-MRI) requires interpretation of four-dimensional DCE data and relies on reader expertise, which could lead to a substantial amount of false positive results.

GBCA is a commonly used intravenous contrast agent for DCE-MRI, which induces an increment in the longitudinal relaxation rate, exhibited as an enhancement in the signal intensity of T1-weighted images. The pharmacokinetic parameters (Ktrans, etc.) could be obtained by fitting an appropriate model to the signal intensity time course of serial acquisition T1-weighted images after an injected GBCA into a tissue of interest (2). Several researches have proven that these parameters changes corresponded to the damage of microstructure in tumor tissues (3–5).

Texture feature is a quite abstract concept for medical imaging, it is an important surface characteristic used to identify and recognize objects, for which it might qualified the ability to describe the internal structures of lesion tissues (6). Texture extracted from breast DCE images enabled rich properties of region of interest (ROI) areas which were also useful for diagnosing different types breast lesions (7).

Texture feature represents the appearance of the medical images and considers how its pixel intensity were distributed (8), it also focuses on the intensity distribution and the relationships of neighboring pixels. Various texture analysis approaches tend to represent the examined textures from different perspectives. Texture analysis was also sensitivity to the subtle changes of the tissues that might ambiguously for human. It has been successfully used in several radiological studies, including subchondral bone, glioma tumor, distinction analysis of invasive adenocarcinoma and pre-invasive or minimally invasive adenocarcinoma (9–11).

Previously researches prefer to explore the variances of texture parameters between different types of diseased tissue. Loose et al (12) evaluated the Haralick textures on DCE-MRI data for the differentiation of Breast Tumor Lesions, Toshiaki et al (13) adopted the histogram analysis to the evaluation of vascular permeability in glioma. On the other hand, chemometric methods such as principal components analysis (PCA), support vector machine (SVM), linear discriminant analysis (LDA), have been employed to the discriminant analysis of different types of lesions (9,14,15), which aims to access the overall changes in multi-texture parameters.

In the present study, pharmacokinetic parameters, as well as the texture parameters that obtained from the DCE-MR imaging and parameter maps were discussed, PCA and Fisher discriminant analysis (FDA) were adopted to differentiate the benign and malignant breast tumors based on the textures, respectively. As to our best knowledge, this is the first try to evaluate the diagnostic efficiency of FDA method on breast tumors.

Materials and methods

Patients data

The present study was approved by the institutional review board of The Wei Fang Traditional Chinese Hospital. From December 2016 to June 2017, 56 patients with an average age of 45.3±11.1 years (25–69 years) were retrospectively enrolled. Informed consent was signed before the patients underwent contrast enhanced MR examination. Pathological examinations were performed to confirm the MR examination results. All patients were sequenced from 1 to 56 according to examination date. Fifty-six patients (23 cancer tumors, 33 benign tumors) were divided into malignant and benign group, and the benign case including breast adenoma and intraductal papilloma.

The patient selection criteria of the present study were as follows: i) All lesions were histopathological confirmed by breast surgery; ii) MR dynamic enhancement was performed no >30 days before surgery or biopsy.

MRI protocol

An MRI scan of the whole breast was performed using a 3T whole-body scanner (GE Discovery MR 750; GE Healthcare, Milwaukee, WI, USA) equipped with a dedicated 8-channel double-breast phased-array coil in prone position. Breathe training was given before MRI scan.

A series of unenhanced MRI scanning sequence were performed first, multi-flip angle (5, 10 and 15°C) T1 Mapping was collected firstly with only one phase, respectively. DCE MR imaging was collected using vibrant sequence with the following parameters: Repetition time/echo time=4.1/2.1 ms, flip angle=10°C, slice thickness=4 mm, matrix size=256×256 pixels, number of slices=120, and the field of view (FOV) was adjusted according to the breast volume. For each patient, a total of 30 acquisitions were obtained.

Thirty measurements with a temporal spacing of ~10 sec. At the beginning of DCE-MRI, a baseline acquisition was collected with two unenhanced phase scans, a bolus of 0.1 mmol/kg gadolinium contrast agent (Omniscan; GE Healthcare, Shanghai, China) was injected intravenously at a rate of 4 ml/sec followed by 20 ml saline flush through antecubital vein at the third acquisition. The scan duration after bolus injection was ~380 sec.

Pharmacokinetic parameters and texture features

The DCE-MRI data were transferred into Omni-Kinetic software (GE Healthcare) to obtain the perfusion and permeability parameters of the lesion areas in breast, using the dual-compartment pharmacokinetic model of Extended Tofts Linear (16,17). This model evaluates Ktrans, Ve, Vp and Kep values by analog the EES space in tissues. Arterial input function (AIF) describes the time-dependent contrast agent input to the ROI (18), Subsequently, the semi-quantitative parameters, such as the slope, maximum enhancement, and area under the curve (AUC) were computed based on the model in software.

ROIs were selected based on enhanced T1-weighted images and then responded to the pharmacokinetic maps. For each breast lesions, 1–3 ROI with 10–30 mm2 size were manually positioned on the T1WI imaging by an experienced radiologist (Fig. 1). In Fig. 1, a benign and a malignant lesion was clearly exhibited (Fig. 1A and B), a circle or a polygon ROI would be drawn manually according to the shapes of lesion.

Figure 1. ROI selection on the T1WI images of patients, a round or polygonal ROI was obtained as shown in (A) and (B). ROI, region of interest

The heterogeneity analysis (i.e. histogram, Gray-level co-occurrence matrix, etc.) was automatically calculated by the Omni-Kinetic software. For each patient, all the texture features were obtained from DCE-MRI images, Ktrans and AUC maps, respectively. Texture features from DCE-MRI images were generated only using arterial phase images according to the AIF curves. All ROIs used to the texture analysis were corresponded to the pharmacokinetic calculation parts.

Different computing methods depicted the ROI pixel distribution from different aspects. Histogram described the pixel distribution, while the Gray-level co-occurrence matrix explained the relationship of adjacent pixels (19). In this study, totally 67 texture features from 4 different calculation methods, were achieved as below: First Order (n=14), Histogram (n=15), Grey level co-occurrence matrix (GLCM, n=13), Haralick matrix (n=9), and grey level run length matrix (GLRLM, n=16).

Statistical analysis

PCA and FDA was carried out in IBM SPSS Statistics software (version 19.0; BM SPSS, Chicago, IL, USA) which is a multivariate data analysis software. PCA method extracted the principal component information of the texture dataset, the first principal component scores could reflect the type of each samples in the patient dataset grossly. FDA is a classical analytical technique for feature extraction and classification, a mapping was first calculated to reduce the dimension of texture dataset, then a linear discriminant analysis was built based on the variance analysis so that several groups or classes could be discriminated as clearly as possible (20–22). One-way analysis of variance (ANOVA) test was performed for the Ktrans and AUC values using the software of Origin 8.0 (OriginLab, Northampton, MA, USA) (23).

Results

DCE-MR images of two patients with breast lesions were shown in Fig. 2, the corresponding pharmacokinetic parameter maps that calculated by the Extended Tofts Linear model were also gave out, including Ktrans images and AUC images. The lesion regions were obviously highlighted in the parameter images, which indicated a more higher permeability and perfusion when compared to the adjacent tissues (24,25).

Figure 2. Ktrans and AUC images derived from the DCE-MRI images by Extended Tofts Linear model. The lesion areas were obviously spotlighted in the parameter maps. AUC, area under the curve; DCE-MRI, dynamic contrast enhanced magnetic resonance imaging.

One-way analysis of variance (ANOVA) test was performed for the Ktrans and AUC values using the software of Origin 8.0. In this study, a significant difference on the Ktrans values was found between the malignant breast tumors and benign breast tumors (P=0.0011, Fig. 3A). However, the ANOVA test on the AUC values showed no significant differences (P=0.29), we could not characterize the benign and malignant tumors only according the averaged AUC values of the two groups samples.

Figure 3. One-way analysis of variance tests on the (A) Ktrans and (B) AUC values for malignant and benign breast tumors. AUC, area under the curve.

Previously studies have revealed the potential diagnostic ability of texture feature from medical images. In this study, 67 texture features were extracted from the DCE-MRI images and the pharmacokinetic parameter maps, including Ktrans and AUC (as shown in Fig. 1), respectively.

PCA was firstly applied to the texture features of the DCE-MRI images, Ktrans and AUC maps, respectively, without any pre-processing. Recently study has tried to use the PCA method followed by adjusted model to analysis breast DCE-MRI images, and an objective diagnostic tool was established (26). However, PCA could also be used as an early attempt to classify unknown datasets, as shown in Fig. 4, the Factor-1 score values of PCA could roughly divide the 56 patients into two groups, the hollow squares (□), circles (○), and triangles (∆) were the misjudged cases (Fig. 4).

Figure 4. The first component scores of PCA on the texture parameters from DCE-MR images (A), Ktrans images (B) and AUC images (C). PCA, principal components analysis.

To achieved the better discriminant accurate results of texture features, further research was required. Spearman correlation analysis was firstly performed to reduce the redundancy of the texture feature on T1WI images, Ktrans and AUC maps. One texture feature would be discard once the Spearman correlation coefficient between two texture features was >0.9, the correlation analysis was carried out for all texture features.

Totally seven common features were left in the three texture feature datasets (Table I). One-way ANOVA tests were calculated for the seven common features on T1WI, Ktrans and AUC maps, respectively. Haralick matrix showed a promising differentiation ability for the malignant and benign tumors, both on the T1WI images and pharmacokinetic parameter maps (P<0.05), on the other hand, the Long-Run-Low Grey-Level-Emphasis of GLRLM exhibit the same ability (6,8).

Table I. ANOVA tests on the texture features from T1WI image, AUC and Ktrans maps.

Table I.

ANOVA tests on the texture features from T1WI image, AUC and Ktrans maps.

Method	Texture feature	T1WI image	Ktrans maps	AUC map
Haralick matrix	Difference entropy	P<0.05	P<0.05	P<0.05
	Inverse difference moment	P<0.05	P<0.05	P<0.05
Grey level run-length matrix	Min intensity	P>0.05	P<0.05	P<0.05
	Max intensity	P>0.05	P>0.05	P>0.05
	Max size	P>0.05	P<0.05	P>0.05
	Long run low grey level emphasis	P<0.05	P<0.05	P<0.05
	Long run high grey level emphasis	P>0.05	P>0.05	P>0.05

[i] ANOVA, one-way analysis of variance; AUC area under the curve.

FDA method was also built based on the three texture features datasets that have underwent the spearman correlation analysis, the discriminant accuracies of FDA method on the DCE-MRI images, Ktrans maps and AUC maps were 93, 98 and 98%, respectively (Table II). Cross validation analysis was performed on the FDA method for DCE-MRI images, Ktrans and AUC maps with the discriminant accuracies of 82, 77 and 77%, respectively. The promising results of FDA and cross validation result indicated that FDA method could be a reliable tool to diagnose the breast lesions.

Table II. FDA method on DCE-MRI images, Ktrans and AUC maps.

Table II.

FDA method on DCE-MRI images, Ktrans and AUC maps.

		Predicted
Imaging	Type	Malignant	Benign	Discriminant accuracy (%)	Discriminant accuracy (%)
DCE-MR images	Malignant	21	2	91	93
	Benign	2	31	94
Ktrans maps	Malignant	23	0	100	98
	Benign	1	32	97
AUC maps	Malignant	23	0	100	98
	Benign	1	32	97

[i] FDA, Fisher discriminant analysis; DCE-MRI, dynamic contrast enhanced magnetic resonance imaging; AUC area under the curve.

Discussion

In this study, a comprehensive exploring, including pharmacokinetic parameters, texture features, and the created FDA method, was adopted to quantitatively evaluate the malignant and benign breast tumors. All these three aspects showed a promising result for the breast patients. This experiment provided a preoperative clinical assessment of breast tumor.

In recent years, pharmacokinetic parameters have been adopted to explore the penetration and perfusion changes inside the tumors (27–29). In the present study, the semi-quantitative parameters (e.g. AUC) and quantitative parameters (e.g. Ktrans) were calculated, Ktrans values showed a better diagnostic ability between the malignant and benign breast tumors than AUC. This difference of Ktrans and AUC values has indicated the increment of microvascular permeability in the malignant tumors. Ktrans is a well-established permeability parameter of tumors, which was obtained with DCE-MRI and has been applied to many kinds of tumors. It reflects the diffusive transport of contrast agent across the capillary endothelium (16,17). In this particular circumstance, Ktrans might directly reveal the microstructure changes inside tumors than AUC values. AUC described the areas under the arterial input function (AIF) that obtained by the Omni-Kinetic software. No significance differences were found between the benign and malignant breast tumors. The AIF curves were slightly different among the patients, which were easily affected by the machine and scanning mode. The AUC values might be unsteadily. The results in this study indicated that more attentions should be gave to the Ktrans values rather than AUC, pharmacokinetic parameters depict tumors based on the relationships between the permeability changes and biological characteristics, which might be have a better expression on the tumors.

Compared to pharmacokinetic parameters, texture features are a quite abstract concept for medical images. It considers the pixel distribution in ROIs through engineering methods. Although lots of researches have refer to these areas, some basic concepts were still ambiguous; texture parameters exhibit excellent skills on the tumors assessment and grade, survival prediction for cancer patients, evaluation of tumor therapy and prediction distant metastasis (30–32). However, how to define the texture features in medical images were still questions that need to be explained urgently. In this study, texture features on DCE-MRI images, Ktrans and AUC maps gave out difference results, For the seven common texture features that underwent spearman correlation analysis, five texture features from Ktrans map exhibit significant difference. On the other hand, difference entropy, inverse difference moment of Haralick matrix and the long-run-low grey-level-emphasis of GLRLM showed a promising ability on the three types of images. All the significant texture features characteristic the breast tumors in different aspects. The research is ongoing in our department to reveal its clinical sense or to connect the texture feature to any related clinical standards.

In addition to explore the diagnostic value of every single texture feature, regression analysis was the common way to find out the capacity of whole texture features (33,34). FDA was adopted in this study, which was easily built without strictly application restrictions. The discriminant accuracies of FDA on the DCE-MRI images, Ktrans and AUC maps were 93, 98 and 98%. A promising cross validation result on the DCE-MRI images, Ktrans and AUC maps were also achieved. The results indicated the feasibility of FDA method on breast tumors. However, some limitations should be acknowledged in this mathematical regression analysis. Firstly, Since the limitation of patient size, further investigation with much larger patient sample sizes should be included to improve the differentiation accuracies and model stability. Secondly, in this study, the patients were only divided into benign and malignant groups, more patients should be involved to specify the disease of benign and malignant breast tumors, an overall analysis on the breast tumors would be cleared once enough patients were given. Thirdly, texture analysis and the regression were generated using arterial phase images, some slight differences might be existed between different phase images, which was not discussed in this study.

The volume data of all breast tumors were obtained on the AW workstation of GE Discovery MR 750, Specified morphological data in tumors, such as surface area, were also derived from the Omini-kinetic software. Since the limitation of the patient amount, tumor morphological description was not regarded as an influential factor in this study. It would be essential and interesting to clarify the pharmacokinetic parameter changes of tumors with different size. On the other hand, metastatic lymph node plays as a significant role for tumors, the prediction of lymph node metastasis has begun to be involved in other fields with excellent results (35). Thus, the investigation of metastatic lymph node in breast cancer could also be feasible by the methods adopted in the research. Related researches were ongoing in our department.

Compared to the previous study that either focused on the analysis of each texture feature or emphasis the importance and potential of regression model, this study explored the pharmacokinetic parameters, texture features and the regression model on breast tumors, a comprehensive analysis were carried out in this study to improve understanding of quantitative parameters on breast tumors.

In conclusion, pharmacokinetic parameters and texture features provide quantitative measures of heterogeneity in breast tumors at different aspect. In this study, all parameters showed a valuable ability in breast lesion discrimination and characterization. On the other hand, the combination of regression model with quantitative parameter also exhibit promising capacity, further exploration was needed to validate the diagnostic ability of pharmacokinetic parameters, texture features and the mathematical regression analysis.

Acknowledgements

Not applicable.

Funding

No funding was received.

Availability of data and materials

The datasets analysis methods used during the current study are available from the corresponding author on reasonable request.

Authors' contributions

QLN was responsible for designing the study and drafting the manuscript. XMJ, QL, ZLZ performed the statistical analysis and researched the literature. HWD and SSW were responsible for the MRI scanning and XXZ contributed to DCE-MRI. All authors read and approved the final version of the manuscript.

Ethics approval and consent to participate

This study was approved by the institutional review board of The Wei Fang Traditional Chinese Hospital. Informed consent was obtained from all patients.

Consent for publication

Not applicable.

Competing interests

The authors declare that they have no competing interests.

References