In total, 320 patients came to Sun Yat-sen University Cancer Center for clinically indicated bilateral breast MRI between September 2008 and December 2011. The patients' breast MRI examinations were retrospectively reviewed. The MRI studies of 205 women with 275 breast lesions met the following inclusion criteria in the present study: i) MRI was performed using a 1.5-T magnet; ii) both DCE-MRI imaging and DWI imaging sequences were performed; iii) diagnosis was confirmed following pathological analysis subsequent to core-needle biopsy or surgical excision (248 lesions), or lesion stability was confirmed at a minimum follow-up of 2 years (27 lesions); iv) lesions presented as a mass according to the breast imaging reporting and data system MRI lexicon; and v) patients had not received a biopsy or received therapy prior to MRI examination. Of the 320 patients, 23 were excluded for not having a suspicious abnormality on dynamic images, 40 were excluded for non-mass-like enhancement, 16 were excluded because their pathological results were not determined, 20 were excluded for having lesions without a sufficient follow-up period, and 16 were excluded for inadequate fat suppression on DWI images.

Imaging was performed with a GE Signa HDx 1.5-T superconductive magnetic system, using a bilateral, dedicated four-channel phased-array breast coil in the prone position. Standard imaging was performed, including an axial fast spin echo (FSE) T1WI and an axial and sagittal FSE T2WI. Subsequently, DWI images were acquired in the axial planes, prior to gadolinium-based contrast material injection, using: A spin-echo single-shot echo planar imaging sequence; array spatial sensitivity encoding technique (acceleration factor of two); b values of 0 and 800 sec/mm²; fat suppression; 5,000/75 (repetition time msec/echo time msec); 5 mm section thickness; a 30×30 cm field of view; a 128×128 matrix; 0 mm section gap; and 130 sec acquisition time.

Subsequently, after one set of unenhanced baseline images, dynamic contrast-enhanced MRI data were acquired using an MRI-specific automatic power injector (Medrad Inc., Pittsburgh, PA, USA) to inject 0.1 mmol/kg body weight contrast medium gadolinium diethylenetriamine penta-acetic acid (Gd-DTPA), with a hand venipuncture technique at a rate of 3 ml/sec. Saline (10 ml at 3 ml/sec) was then injected to wash the tube. Dynamic scanning was initiated by simultaneously pushing the high-pressure syringe button and the dynamic scan button. Nine post-contrast sets were acquired. Each sequence was performed in the sagittal plane at 20 sec intervals with fat suppression using: Three-dimensional spoiled gradient recall echo sequence; 5.5/2.6 repetition time msec/echo time msec; 3.4 mm slice thickness; 15° flip angle; a 22×22 cm field of view; a 288×192 matrix; and a 59 sec acquisition time. Subsequently, axial MRI was employed using fat-suppressed enhanced T1WI sequence.

All images were analyzed independently by two radiologists with ten years experience in interpreting breast MRI. They were blinded to the histological results of current patients. The images were assessed independently and any disagreements were resolved by achieving consensus. All lesions were assessed using The Breast Imaging Reporting and Data System (BI-RADS) (15). BI-RADS category 1 (negative) and category 2 (benign) denote an essentially 0% likelihood of cancer. BI-RADS category 3 (probably benign) assessment is more intuitive and can be recommended in the case of a unique focal finding for which the likelihood of malignancy is ≥0% but ≤2%. BI-RADS category 4 (suspicious) and category 5 (highly suggestive of malignancy) describe MRI findings that are suspicious enough to warrant tissue diagnosis. BI-RADS category 6 (known biopsy-proven malignancy) describes MRI findings of biopsy-proven breast cancer for which surgical excision is recommended when clinically appropriate. All images were analyzed on a workstation (Centricity Radiology RA 600 V 7.0; GE Healthcare, Chicago, IL, USA). All quantitative analysis software was written in MATLAB (MathWorks, Natick, MA, USA; http//www.mathworks.com).

For kinetic analysis, time-signal intensity curves were obtained from manually drawn regions of interest (ROIs), the size of which varied with the size of the enhancing lesion and were chosen to selectively include the area with the strongest enhancement, as identified on the first post contrast subtracted image. The early-phase enhancement rate and the signal enhancement ratio (SER) (16) were quantified by means of an ROI-based determination of lesion signal intensity prior and subsequent to the injection of Gd-DTPA. The early-phase enhancement was calculated according to the enhancement formula (SI₁-SI₀)/SI₀ ×100 (17) and SER was defined as (SI₁-SI₀)/(SI_last-SI₀), where SI₀, SI₁, and SI_last represent the signal intensity in the pre-contrast, the first post-contrast and the last images, respectively.

Following manual lesion identification, automatic segmentation was undertaken for lesion contours. The segmentation used a novel two-step approach that incorporated fuzzy c-means (FCM) clustering (18) and gradient vector flow snake algorithm (19). The FCM clustering based method was used to obtain initialization for commencement of segmentation while the gradient vector flow snake model was further applied to obtain the exact segmentation (Fig. 1). A total of 13 texture features were estimated from the gray level co-occurrence matrix (GLCM) (20). The parameters included: Angular second moment; contrast; correlation; inverse difference moment; sum average; sum variance; sum entropy; entropy; difference average; difference variance; difference entropy; information measure of correlation 1; and information measure of correlation 2 (21). In total, 11 morphological features, including compactness, spiculation, extent, elongation, solidity, circularity, entropy of radial length distribution, fractal, heterogeneity, area and eccentricity, were also computed (22).

Breast masses in the study were all verified histopathologically (Table I), or the diagnosis was confirmed following at ≥2 years of follow-up. Lesion status was used as baseline for statistical evaluation of the performance of the features. The whole patients set was randomly divided into two sets, one for classifier training and the other for testing the performance of the classifier. The ten-fold cross validation scheme was used to evaluate and find the classifier with the best performance.

The ROIs were manually drawn on the diffusion-weighted images (b=800 s/mm²; Figs. 2 and 3) and were placed in regions with high signal intensity on the images. The contrast and morphological characteristics at the early phase of contrast-enhanced T1-weighted imaging and T2-weighted imaging were used to guide ROI placement to prevent areas of T2 shine-through, which was usually found in necrotic or cystic parts. The ROIs were defined as the area slightly smaller than the actual lesions to reduce partial volume effects. In the current study, all ROIs were >20 mm² (23). The DWI intensity for each lesion was classified as high or low compared to that of the corresponding background breast tissue. The mean ADC of the voxels in ROIs were then obtained.

Morphological features were shown to perform well in terms of discriminating lesion types (8). Meanwhile, using the textural or kinetic measurements alone can also shade light in the characteristics of lesions, shown in Table II. Therefore, the combination of the merits of the characterization parameters, including morphological, textural and kinetic measurements, was expected to achieve a better, or at least comparable performance compared with individual measurements. The combination of the parameters as whole reflected different aspects of lesion properties and was a potentially comprehensive approach to the characterization of lesion status.

In total, 24 features were initially estimated from the segmented lesion image. Since the features may be self-dependent or irrelevant to the lesion type, an exhaustive feature selection method was used to find an optimal feature subset in which the classification accuracy is highest (24). The differentiation of malignant from benign lesions was treated as a two-class pattern classification problem. A newly reported classification model, the difference-weighted local hyperplane classification model (DWLH), was used to select the feature subset and to evaluate the performance of the combined feature subsets (25). In the classification model, a local hyperplane is constructed from its nearest neighbor for each observed sample. The label of the sample is decided by the minimized distance to its class-dependent hyperplane. The performance of the model is superior to that of the classical k-nearest neighbors algorithm in that the local hyperplane is robust to noises and outliers (26). The present study selected DWLH, not only because of its superior performance over classical models, including the support vector machine, but also as a result of the rationale behind its mathematical formulation (27). Since lesions usually share certain pictorial similarities among patients with the same type lesions, the status of lesion could be deducted from its nearest neighbors, other than particular one as k-nearest neighbors does. This follows the same rationale with the classifier of DWLH.

To test the discrimination power of different characterization of breast masses, the whole dataset where categorized into six subgroups including morphology, kinetics, texture features and their combinations. To further evaluate the discrimination power of ADC, various characteristics of breast masses were combined with/without ADC, resulting another three data subgroups. On each subgroup, the classifier of DWLH was evaluated via ten-fold cross validation scheme. For each experiment, the averaged accuracy and AUC were calculated to serve as criteria of the feature performance. To eliminate the statistical variations during the training phase, the present study conducted 10 classification experiments independently on each dataset. The averaged classification error was recorded and is presented in Table III. Receiver operating characteristic (ROC) analysis was also used to evaluate the diagnostic performance of the model under various features. The area under the ROC curve (AUC) served as the criterion for selecting the best combination.

Segmentation results of randomly selected sample breast lesion can be seen in Fig. 1. In total, 6 features performed optimally in terms of prediction. The features included two shape (heterogeneity, area) and three texture (sum variance, sum entropy, difference entropy) parameters and one kinetic parameter (SER).

The diagnostic performance of each feature used in the experiment classifier was also evaluated individually. The mean and standard deviation and the diagnostic performance of the 6 selected parameters are summarized in Table II. Among the 6 parameters, the diagnostic accuracy of SER was the highest.

Utilization of 5 anatomical features, including texture and shape parameters, resulted in an AUC of 0.68. The kinetic parameter SER reached a comparable AUC of 0.84. When combining these 6 parameters into a unified diagnostic model, the AUC was further improved to 0.86. This combined classifier achieved a sensitivity of 0.81 and a specificity of 0.91 with a diagnostic accuracy of 0.87 (Table III).

A malignant lesion and a benign lesion can be seen in Fig. 2 and in Fig. 3, respectively. The box plot of the ADC for all patients is shown in Fig. 4. An overlap of ADC was observed between malignant and benign breast lesions. The mean ADC of malignant lesions (mean, 1.16×10⁻³ mm²/sec; 95% CI, 1.13×10⁻³ mm²/sec; 1.20×10⁻³ mm²/sec) was significantly lower than that observed for benign lesions (P<0.05; mean, 1.68×10⁻³ mm²/sec; 95% CI, 1.62×10⁻³ mm²/sec; 1.75×10⁻³ mm²/sec). The predictive accuracy (sensitivity, specificity) of mean ADC was 0.84 (0.78 and 0.90, respectively; Table III). Thus, a threshold value of 1.37×10⁻³ mm²/sec can achieve a specificity of 0.90, with a sensitivity of 0.78, to discriminate between malignant and benign lesions.

When the ADC value was incorporated into the feature sets, the sensitivity, specificity and accuracy of the classification model increased to 0.85, 0.94 and 0.90, respectively (Table III). The value of AUC was dramatically increased from 0.86 to 0.90 by incorporating without/with using ADC (Fig. 5). It implies that the ADC possesses the potential power in discriminating benign masses from malignant ones. It also suggests that addition of ADC values enhances the discriminatory power of the combined feature set.