Applying a New Quantitative Image Analysis Scheme based on Global Mammographic Features to Assist Diagnosis of Breast Cancer

Abstract

Background and Objective:

This study aims to develop and evaluate a unique global mammographic image feature analysis scheme to predict likelihood of a case depicting the detected suspicious breast mass being malignant for breast cancer.

Methods:

From the entire breast area depicting on the mammograms, 59 features were initially computed to characterize the breast tissue properties at both spatial and frequency domain. Given that each case consists of 2 cranio-caudal and 2 medio-lateral oblique view images of left and right breasts, two feature pools were built, which contain the computed features from either 2 positive images of one breast or all the 4 images of two breasts. Next, for each feature pool, a particle swarm optimization (PSO) method was applied to determine the optimal feature cluster followed by training a support vector machine (SVM) classifier to generate a final score for predicting likelihood of the case being malignant. To test the scheme, we assembled a dataset involving 275 patients who had biopsy due to the suspicious findings on mammograms. Among them, 134 are malignant and 141 are benign. A ten-fold cross validation method was used to train and test the scheme.

Results:

The classification performance levels measured by the areas under ROC curves are 0.79±0.07 and 0.75±0.08 when applying the SVM classifiers trained using image features computed from 2 and 4 view images, respectively.

Conclusions:

This study demonstrates feasibility of developing a new global mammographic image feature analysis-based scheme to predict the likelihood of case being malignant without lesion segmentation.

1. Introduction

Since the breast lesions are highly heterogeneous containing the overlapped dense fibro-glandular tissues, reading and interpreting mammograms is a difficult task for radiologists [1, 2]. Accordingly, developing computer-aided detection and diagnosis (CAD) schemes of mammograms have attracted extensive research interest in the recent decades, which aims to provide radiologists a “second opinion” supporting tool in reading and interpreting mammograms [3]. Currently, there are two types of CAD schemes namely, computer-aided detection (CADe) schemes and computer-aided diagnosis (CADx) schemes. The former detects suspicious lesions and determines their locations in mammograms [4], while in contrast the latter makes classification between malignant and benign lesions [5]. Although commercialized CADe systems are currently available and used in the clinical practice, it is in controversy of whether using these schemes can actually improve radiologists’ performance in detecting breast cancer [6]. Despite the somehow disappointment when using CAD schemes of mammograms in the clinical practice, many researchers believed that we need to continue exploring new approaches to improve CAD performance and optimize CAD in the clinical practice [7].

The performance of CAD schemes for mammograms heavily depends on case difficulty, including the conspicuity of lesions (e.g., fuzziness of lesion boundary) and overlap of dense and heterogeneous fibro-glandular tissues [8]. These difficulties can substantially reduce the accuracy and robustness of lesion segmentation [9], which will affect the performance and reproducibility of CAD schemes [8, 10]. Previous studies demonstrated that improving lesion segmentation accuracy can produce more reliable image features, which helps achieve better CAD performance in classifying between malignant and benign breast lesions [11, 12]. Nonetheless, accurate lesion segmentation is still challenging for digital mammogram processing due to the high heterogeneity of lesions and overlapping of dense fibro-glandular breast tissues on mammograms [1, 13]. In addition, due to the lack of ground truth, evaluating lesion segmentation accuracy is also very difficult or subjective with large inter-reader variability. Thus, in order to avoid this challenge, different approaches have been proposed to develop new CAD schemes without lesion segmentation [12] including using deep learning algorithms [14]. However, for the deep learning based methods, training a reliable deep learning based CAD schemes usually requires a large and diverse image dataset (e.g., using a dataset with 128,175 retinal images [15]), which is often unavailable in cancer imaging field.

In order to overcome these limitations, our recent studies indicate that quantitative image markers extracted from the whole breast areas of mammograms (namely, the global features) can be used to predict short-term breast cancer risk with significantly higher prediction power [16, 17]. Thus, we hypothesized in this study that it is possible to identify and fuse the global image features computed from the whole breast area depicting on mammograms without lesion segmentation, which enables to generate a new quantitative imaging marker for predicting the likelihood of a testing case being malignant. This new global mammographic image feature-based approach cannot only avoid lesion segmentation, but also reduce the requirement of large training dataset as the conventional deep learning approach [16]. Thus. the objective of this study is to develop a new global mammographic image feature analysis-based CAD scheme and validate our study hypothesis. The experimental details are presented as follows.

2. Materials and Methods

2.1. Image Dataset

From the IRB-approved retrospective study protocols, we have assembled a digital mammography image database as reported in our previous studies (e.g., [4, 10, 11]). From the assembled image database, we selected an image dataset for this study, which consists of fully anonymized digital mammograms acquired from 275 women participants in breast cancer screening. Each case has one suspicious mass-type lesion identified and detected by the radiologists in original mammogram reading and interpretation. All suspicious lesions were biopsied and confirmed by histopathology examinations. Among these cases, 134 were confirmed to be malignant; while the other 141 cases were benign. In addition, cancer was detected only in one breast in this dataset.

All digital mammography screening examinations were performed using Hologic Selenia (Hologic Inc) full-field digital mammography (FFDM) systems. Each mammography screening case has 4 images including 2 cranio-caudal (CC) and 2 medio-lateral oblique (MLO) view images of left and right breasts. Since this study only focused on the cases depicting soft tissue mass type lesions, the mammograms were subsampled to reduce the image size, which is a common practice used in CAD research field including the commercialized CAD schemes [17]. Specifically, as reported in our previous computerized scheme [4], the original FFDM images with a pixel size of 0.07mm were pre-subsampled using the average pixel value computed from a 5 × 5 scanning window. Thus, the actual pixel size used in the subsampled image is 0.35mm.

As summarized in table 1, the mammographic density information of malignant and benign cases was identified by radiologists according to BIRADS guidelines, which shows no significant difference between two classes of malignant and benign cases using the BIRADS based mammographic density ratings. Figure 1 illustrates the example images of one malignant case and one benign case, each of which contains with 4 CC and MLO view images of the left and right breasts. Lesions of the example cases exhibit low conspicuity and high fuzziness, making it difficult for lesion segmentation and risk prediction.

Table 1.

Distribution of mammographic density (BIRADS ratings) for two groups of the cases in dataset.

Characteristic	Malignant cases	Benign cases
Fatty tissue (1)	6	7
Scattered (2)	55	57
Heterogenous (3)	70	73
Extremely dense (4)	3	4

Figure 1.

Examples of malignant and benign cases with four view display. a) Malignant case. b) Benign case.

2.2. A New CAD Scheme

Our proposed CAD scheme was developed in the following 3 steps namely, feature computation, feature selection and case classification. We first built an initial feature pool containing four different groups of features. Next, a particle swarm optimization (PSO) algorithm was applied to select optimal features so that redundant features can be removed from the feature pool. Finally, a popular machine learning classifier namely support vector machine (SVM) was used to predict the risk or likelihood of a case being malignant. The CAD scheme was implemented in MATLAB software environment.

2.2.1. Feature Computation

As shown in figure 2, we used four images (namely, RCC, LCC, RMLO, and LMLO) of CC and MLO view at the left (L) side and right (R) side. Before building feature pools, our CAD scheme first segmented the whole breast area out in each image by removing all possible artifacts or markers outside the breast areas (assigning all background pixels to 0 as shown in black color in figure 2). The computed image features from the segmented whole breast area can be categorized into four groups. The first group includes 17 statistical image features describing breast area shape and density distribution (as shown in Table 2). These statistical features are widely used to quantify pixel value distribution and its heterogeneity in the 2D image [18, 19]. The other three feature groups are block-based Fast Fourier Transform (FFT) features, Discrete Cosine Transform (DCT) features and Wavelet Transform (WT) features, which aim to help detect and analyze local breast tissue distributions in an image. For this purpose, each target image was first divided into blocks with a scanning window size of 8×8 or 9×9, which is determined based on our previous investigation [20]. Then, FFT, DCT and WT were applied on these blocks to construct three feature matrices, in each of which 14 features were computed (Table 2). More details about feature extraction can be found in the supplemental materials. Finally, three groups of FFT, DCT and WT based features (3×14=42 features) and Shape&Density group (17 features) were combined, so that CAD scheme computed 59 initial features from each image.

Figure 2.

Feature extraction considering four view images of breast area

Table 2.

List of four feature groups

Feature Group	Description
Shape&Density	Mean, Std, Convexity, MeanGradient, StdGradient, Skewness, Kurtosis, Energy, Entropy, Max, Min, Median, Range, RMS, MeanDeviation, Uniformity, Correlation
Block-based features (FFT, DCT and WT groups)	Mean, Std, Max, Min, Median, Range, RMS, Energy, Entropy, Skewness, Kurtosis, MeanDeviation, Uniformity, Correlation

Next, two feature pools were built with different view images. The first one used all the four images (namely, RCC, LCC, RMLO, and LMLO) of CC and MLO view at the left (L) breast and right (R) breast (as shown in figure 2). In each case, 59 previously mentioned image features were computed separately from each of the four view images. We can organize each 59 features into a separate vector, so that each case has four feature vectors. A final vector (e.g. feature pool) was generated by adding corresponding or matched features computed from 4 view images together. Such feature generating process is demonstrated in figure 2. Similar to computing the first feature vector with 4 view images, the second pool includes feature vectors that were computed using only 2 positive view images (e.g., LCC and LMLO view images of one breast). Two images of negative breast were ignored. Finally, these two feature pools and vectors were applied as input to train two machine learning classifiers embedded with the feature selection algorithm separately.

2.2.2. Machine learning classifier and performance assessment

Using the created feature vector, a machine learning classifier is applied to generate the optimal feature cluster and predict the likelihood of the case being malignant. Although many different machine learning classifiers can be used for this purpose [21], SVM classifier uses a constructive machine learning process based on the statistical learning theory to minimize the generalization error [22], which is considered a quite robust classifier applied to the relatively small training datasets and has been used in many biomedical engineering applications [23, 24]. As a result, we adopted SVM classifier in our application, which is built based on the SVM tool box under MATLAB environment.

Based on our dataset, we trained and tested the SVM classifier embedded with the PSO feature selection algorithm to minimize the potential bias during feature selection and lesion classification. In addition, the 10-fold cross validation was performed. Specifically, our dataset was randomly divided into 10-folds. As illustrated in figure 3, SVM was trained with 9 folds of data and tested with the remaining one fold. The process was repeated 10 times. In each repetition, the training dataset was used to identify an optimal feature vector using PSO algorithm, which is detailed in the supplemental materials. The following objective function is adopted to control the training outcome [20, 25]:

$$F_{obj}=\alpha [1-\mathit{Mean}(AUC)]+\beta \mathit{Std}(AUC)+\gamma (\mathrm{number}\mathrm{of}\mathrm{selected}\mathrm{features})$$ (1)

Figure 3.

Flowchart of proposed 10-fold cross-validation based training and testing method

In the above Equation, the parameters α, β, γ are weighted coefficients determined based on the feature distribution in the Euclidean space, and AUC is the computed area under a receiver operating characteristics (ROC) curve [26]. When the objective function reaches its minimum, the training process is finished. Then the trained classifier is applied to make prediction for individual case in the testing fold. Thus, through the 10 training and testing iteration cycles, each of 275 cases in our dataset will be independently tested once and receive a classification score indicating the likelihood of the case being malignant. Finally, based on the classification scores of all 275 cases, the performance of the proposed CAD scheme will be evaluated and compared using AUC and other evaluation indices (e.g. classification sensitivity, specificity, positive and negative predictive values).

3. Experiments and Results

3.1. Evaluation of single features

We computed the Pearson correlation coefficients of all 59 initially computed features with 275 observations associated with each feature. Given either two or four view images of each case, the value of correlation coefficients falls into eight categories as shown in figure 4. The charts show that more than 70% of the absolute correlation coefficients were less than 0.4, which indicates that the feature pool designed in our study provided a comprehensive view of the cases with a relatively small redundancy.

Figure 4.

Feature correlation coefficients analysis of using a) two marked view images and b) four view images.

Figure 5 shows the sorted results of the areas under ROC curves (AUC values) computed from all 59 individual image features. When using features computed from four view images, the top three best performed features including Energy_FFT, Energy_DCT, and Mean_Density with an AUC values of 0.689±0.041, 0.668±0.042 and 0.667±0.042, respectively. Similarly, among all features computed from two positive view images, the top three features are MeanGradient, MeanDeviation_DCT, and Mean_FFT with the AUC value of 0.678±0.042, 0.668±0.042 and 0.665±0.042, respectively.

Figure 5.

The graph of sorting list of AUC values of 59 individual features

3.2. Performance Assessment of the SVM Classifiers

Using classification scores generated by the SVM classifiers on the total 275 cases, we conducted data analysis to assess the SVM classifier’s performance with respect to the training dataset and test dataset. When using the training dataset, figure 6 (a) shows two ROC curves for 2 and 4 view images, which yields AUC values of 0.81±0.036 and 0.79±0.038 (p = 0.057), respectively. On the testing dataset (figure 6 (b)), the AUC values are 0.79±0.07 and 0.75±0.08 for the 2 and 4 view images (p < 0.01). The comparable AUC values between training and testing results indicated that the SVM classifiers were not over-trained and quite robust.

Figure 6.

Comparison of two ROC curves generated using (a) training dataset and (b) testing dataset.

After applying an operating threshold (T = 0.5) to divide the test cases into two predicted classes being malignant or benign, we generated two confusion matrices of testing results, which are presented at Table 3. Based on the two confusion matrices, we further computed other assessment indices such as classification sensitivity, specificity, positive and negative predictive values, and odds ratio, which were summarized in Table 4. The results show that the SVM classifier optimized with features computed from 2 positive view images yielded better classification performance than the classifier optimized with 4 view image features.

Table 3.

Two confusion matrices of applying two SVMs to classify between malignant and benign cases.

	Two marked view images		Four view images
Prediction╲Actual	Malignant	Benign	Malignant	Benign
Malignant	103	27	102	37
Benign	31	114	32	104

Table 4.

Summary of other assessment indices of two SVM classifiers optimized using 2 and 4 view images.

	Sensitivity	Specificity	PPV	NPV	Odds Ratio
Two marked view	81%	77%	79%	79%	14.02
Four view	74%	76%	76%	73%	8.96

4. Discussion

Although several imaging modalities have been tested and/or applied as breast cancer screening tools [27], mammography remains the most cost-effective and widely used tool for the population based breast cancer screening. However, a large amount of suspicious breast lesions (particularly the soft tissue masses) can be detected in mammograms and the majority of them are benign. Thus, exploring new approach to develop more effective CAD schemes to assist the classification between malignant and benign cases or lesions depicting on mammograms is crucial to improve the efficacy of the breast cancer screening and diagnosis [7]. In this study, we developed a novel CAD scheme utilizing the global mammographic image features to predict likelihood of the testing cases being malignant without lesion segmentation. As compared to the previously reported research efforts, our investigation has a number of unique characteristics or new observations as follows:

First, instead of computing image features from the segmented lesion and its surrounding area, the new CAD scheme extracts and computes the global image features from the whole breast areas of mammograms in this study. The majority of previous studies of CAD-based breast lesion classification [28] computed image features from the segmented lesions or its neighbors, which may have advantages and disadvantages. The advantages include enabling to compute features that are more focused and/or relevant to the specific lesions, while the disadvantages include the variable or lower accuracy and/or reproducibility of computing the features due to the difficulty and errors in lesion segmentation. In our investigation, when we use the global features extracted from the two positive images of one breast with lesion detected, the trained SVM classifier yielded a comparable performance (i.e. AUC 0.79±0.07) in classification between the malignant and benign cases (i.e., comparing to a previous study that used a similar dataset [11, 23]). The result indicates that the clinically meaningful information is not only focused on the lesion, but also distributes on the entire breast area of mammogram image. In addition, although CAD schemes without lesion segmentation have been previously developed and reported in the literature (i.e., [10, 14]), these schemes computed image features from a fixed region of interest (ROI) covering the suspicious lesions, which have disadvantages or difficult to adaptively identify the optimal size of the ROIs to cover the lesions with varying size and shape. The approach in this study is different. Thus, to the best of our knowledge, this is the first study that investigate the feasibility of developing a global breast image feature-based CAD scheme to classify between malignant and benign mammographic cases, which avoid difficulty in both segmentation of the lesions and determination of the optimal ROIs, which are the two popular approached used in previous studies.

Second, we trained and tested two SVM classifiers using two feature pools containing the global images features computed from 2 view images of one positive breast and 4 view images of two breasts. The testing results show that the SVM classifier yielded AUC of 0.79±0.07 when 2 view images of one positive breast were involved in the training and testing process. However, when using the image features computed from 4 view images of two breasts to build the SVM classifier, the scheme yielded a reduced performance with an AUC of 0.75±0.08, which implies that the discriminatory information or power may be diluted when adding two negative images of one cancer-free breast. Thus, it should be better to use two view images of one breast to train the SVM classifier. Then, CAD scheme can be applied to two view images of left and right breasts separately. The higher classification score should be selected to represent the likelihood of the testing case being malignant.

Third, unlike many previously developed CAD schemes that focus on computing the morphological and density distribution based features in the spatial domain, we computed image features in both spatial domain (Shape&Density group) and frequency domain (FFT, DCT, Wavelet block-based groups). By evaluating and comparing the prediction performance of every single feature, we observed that the top three features computed from 4 view images of two breasts were Energy_FFT, Energy_DCT and Mean_Density; and the top three features computed from 2 view images of positive breasts were MeanGradient, MeanDeviation_DCT, and Mean_FFT. This difference may be due to the nature of the two view and four view images. As verified in this study, the normal tissues on the mammogram also contain clinically descriptive information for mass classification. However, the normal and abnormal tissues depict significantly different properties on the mammograms, thus different types of features are needed to identify and collect the relevantly useful information from both normal and abnormal the tissue structures. Since two view images contain positive masses, the top features should have a balanced capability to collect the discriminative information from both the masses and the normal tissues. For the four view images, given that two images are completely normal, the selected features should have a better capability to extract the discriminant characters from the normal tissues. In addition, the feature analysis results also show that the copious lesion pattern information exists in both spatial and frequency domain, which was indicted in our previous investigation of assessing response of the metastatic tumors to chemotherapy using CT images [29] and verified in this study for classification between malignant and benign mammographic image cases.

Fourth, since identifying optimal and non-redundant images features is one of the most important and challenged tasks in developing the conventional machine learning classifiers including the SVM classifier [22], we in this study investigated advantages of applying a PSO method to select optimal features and guide the process of training SVM classifier. The results demonstrated that this method enabled to identify and combine the useful pattern inside the global mammographic images features while removing the redundancy. Thus, using this optimal feature selection method, we are able to use a relatively small dataset of 275 cases for the SVM model training and optimization, which avoids the large database requirement when developing the deep learning based CAD schemes.

Despite the encouraging results, we recognize that this study has the following limitations. First, the database established in this investigation was only from one institution with a limited number of cases; therefore, a more diversified dataset including the cases from more than one institution would be desirable to further test the reliability and robustness of our proposed scheme. Second, the initial feature pool with 59 features used in this study may not be an optimal feature pool. We should expand the feature pool using a list of functional and diversified features summarized in the literature [30]. Meanwhile, it is also worth investigating different feature selection methods such as Relief [31], recursive feature elimination [32], variable ranking techniques [33], supervised training [34] or combining them with our PSO-SVM strategy. Third, this preliminary study used the supervised SVM classifiers that have strengths of solving complex problems and adapting well to high dimensional data (or feature vector); however, there may exist necessities to explore other effective classifiers (e.g. linear discriminant analysis (LDA) and artificial neural networks (ANNs)), in particular the fusion of multiple classifiers to loosen the data size requirement [35] and balance the computational efforts to resolve the uncertainty of the model [29]. Last, similar to our previous study, which demonstrated that fusion of the complementary information of global (case-based) and regional (lesion-based) mammographic image features had potential to significantly improve CAD performance in detecting suspicious lesions without increase of false-positive rates [34], we will investigate the optimal fusion method to fuse the classification results of this global feature based scheme with the lesion-based scheme [35] to more accurately and robustly classify between malignant and benign mammographic cases in the future.