Keywords

1 Introduction

Breast cancer is the second leading cause of death among women worldwide [1]. In 2019, 268,600 new cases of invasive breast cancer were expected to be diagnosed in women in the U.S., along with 62,930 new cases of non-invasive breast cancer [2]. Early detection is the best way to increase the chance of treatment and survivability. Data mining has become a popular tool for knowledge discovery which shows good results in marketing, social science, finance and medicine [19, 20]. Recently, multiple classifiers algorithms are applied on medical datasets to perform predictive analysis about patients and their medical diagnosis [6, 9, 10, 21]. For example, using machine learning techniques to assess tumor behavior for breast cancer patients. One problem is that there is a class imbalance in the training data, since the probability of not having this disease is higher than the one of having it. This paper introduces a comparison between three different classifiers: J48, NB, and SMO with respect to accuracy in detection of breast cancer. Our aim is to prepare the dataset by proposing a suitable method that can manage the imbalanced dataset and the missing values, to enhance the classifier’s performance. All tasks were conducted using Weka 3.8.3.

The remainder of this paper is organized as follows. Section 2 presents literature review. Section 3 introduces the datasets. Section 4 describes the research methodology including pre-processing experiments, classification and performance evaluation criteria. The experimental results are presented in Sect. 5. Finally, Sect. 6 shows the conclusion and future work.

2 Literature Review

In recent years, several studies have applied data mining algorithms on different medical datasets to classify Breast Cancer. These algorithms show good classification results and encourage many researchers to apply these kind of algorithms to solve challenging tasks. In [21], a convolutional neural network (CNN) was used to predict and classify the invasive ductal carcinoma in breast histology images with an accuracy of almost 88%. Moreover, data mining is used widely in medical fields to predict and classify abnormal events to create a better understanding of any incurable diseases such as cancer. The outcomes of using data mining in classification are promising for breast cancer detection. Therefore, data mining approach is used in this work. A list of some literature studies related to this method is presented in Table 1.

Table 1. Breast cancer detection research using different machine learning algorithms.
Full size table

3 Datasets

The datasets that are used in this paper are available at the UCI Machine Learning Repository [13].

3.1 WBC Dataset

The WBC dataset contains 699 instances and 11 attributes in which 458 were benign and 241 were malignant cases [14]. In the WBC, the value of the attribute (Bare Nuclei) status was missing for 16 records. Hence data preprocessing is essential and important for this dataset, requiring us to manage the imbalanced data and the missing values.

3.2 Breast Cancer Dataset

The feature form this dataset are computed from a digitized image of a fine needle aspirate (FNA) of a breast tumor. The target feature records the prognosis (i.e., malignant or benign). The dataset contains 286 instances and 10 attributes in which 201 were no-recurrence-events and 85 were recurrence events. In the Breast Cancer dataset, the value of the attribute (node-caps) status was missing in 8 records.

4 Research Methodology

The two datasets used in this work are vulnerable to missing and imbalanced data therefore, before performing the experiments, a large fraction of this work will be for preprocessing the data in order to enhance the classifier’s performance. Preprocessing will focus on managing the missing values and the imbalanced data. To manage the missing attributes, all the instances with missing values are removed. The imbalance data problem needs to adjust either the classifier or the training set balance. To do so, the resample filter is used to rebalance the data artificially. Then, 10 fold cross validation is applied and finally a comparison between these three classifiers is implemented.

Fig. 1.
figure 1

Proposed breast cancer detection model using Breast Cancer and WBC datasets.

Full size image

4.1 Preprocessing Phase

First, the data were discretized using discretize filter, then missing values were removed from the dataset. Second, instances were resampled using the resample filter in order to maintain the class distribution in the subsample and to bias the class distribution toward a uniform distribution. Section 5 will show that this idea is improving the classifier’s performance. Third, 10 fold cross validation was applied then experiments were carried out over three classifiers Naïve Bayes, SMO and J48, as illustrated in Fig. 1.

In Fig. 1, the data preprocessing technique has been applied including three steps: discretization, instances resampling and removing the missing values. After that, 10 fold cross validation has been applied. Then, three classifiers have been evaluated over the prepared datasets.

4.2 Training and Classification

In order to minimize the bias associated with the random sampling of the training data, we use 10 fold cross validation after the pre-processing phase. In k-fold cross-validation, the original dataset is randomly partitioned into k equal size subsets. The classification model is trained and tested k times. Each time, a single subset is retained as the validation data for testing the model, and the remaining k−1 subsets are used as training data. Three classification techniques were selected: a Naïve Bayes (NB), a Decision Tree built on the J48 algorithm, and a Sequential Minimal Optimization (SMO). The NB classifier is a probabilistic classifier based on the Bayes rule. It works by estimating the portability of each class value that a given instance belongs to that class [15]. The J48 algorithm [16] uses the concept of information entropy and works by splitting each data attributers into smaller datasets in order to examine entropy differences. It is an improved and enhanced version of C4.5 [17]. The SMO model implements John Platt’s sequential minimal optimization algorithm for training a support vector classifiers. This implementation globally replaces all missing values and transforms nominal attributes into binary ones. It also normalizes all attributes by default [18].

4.3 Performance Evaluation Criteria

In this study, we use five performance measures to evaluate all the classifiers: true positive, false positive, ROC curve, standard deviation (Std) and accuracy (AC).

$$ AC = \left( {TP + TN} \right)/\left( {TP + TN + FP + FN} \right). $$
(1)

Where TP, TN, FP and FN denote true positive, true negative, false positive and false negative, respectively.

5 Experimental Results

First, the three classifications algorithms were tested on the WBC and the Breast Cancer datasets without applying the preprocessing techniques. Among them, the best result was recorded for J48: 75.52% in the Breast Cancer dataset and for SMO: 96.99% in the WBC dataset. Next, after applying preprocessing techniques accuracy increases to 98.20% with J48 in the Breast Cancer dataset and 99.56% with SMO in the WBC dataset.

5.1 Experiment Using the Breast Cancer Dataset

First, the three classifiers are tested over original data (without any preprocessing).The results show that J48 is the best one with 75.52% accuracy where the accuracy of NB and SMO are 71.67% and 69.58%, respectively. Next, we apply discretization filter and remove the records with missing values, results improved with NB and SMO as follows: NB: 75.53% and SMO: 72.66% where J48: 74.82%. After that, resample filter was applied for 7 times. The Performance of the classifiers are improved and enhanced as shown in Table 2.

Table 2. Performance of the classifiers in the Breast Cancer Dataset.
Full size table

As illustrated in Table 2, we can obviously notice that the more resample filter we apply, the improved accuracy we obtain. That is because the data is imbalanced and the filter maintains the class distribution. For the Breast cancer dataset, J48 outperforms others with 98.20%. Accuracy measures for J48 classifier is shown in Table 3 and Roc curve of J48 is shown in Fig. 2.

Table 3. Accuracy measures for J48 in the Breast Cancer Dataset.
Full size table
Fig. 2.
figure 2

J48 ROC curve in Breast Cancer Dataset.

Full size image

To measure the performance of the proposed model, we compare the obtained results with the study proposed in [9]. The same dataset and three classifiers including J48 algorithm are used to evaluate the model’s performance. According to the results, the J48 classifier of the proposed model achieves high accuracy comparing to other classifiers. This is because of using the resample filter for the pre-processing phase in the proposed model rather than feature selection technique that used in [9] as illustrated in Table 4.

Table 4. Compression of accuracy measures for the Breast Cancer Dataset.
Full size table

5.2 Experiment Using the WBC Dataset

Same experiments were applied with the WBC dataset. With respect to applying preprocessing techniques all algorithms present higher classification accuracy, the difference lies in the fact that using the resample filter several times improves the classification accuracy. SMO classifier achieve 99.56% efficiency compared to 99.12% of the Naïve Bayes and 99.24% of the J48. Results are illustrated in Table 5.

Table 5. Performance of the classifiers in WBC dataset.
Full size table

In the WBC dataset, SMO superior than others with 99.56%. Accuracy measures for SMO classifier is shown in Table 6 and Roc curve of SMO is shown in Fig. 3.

Table 6. Accuracy measures for SMO in WBC Dataset.
Full size table
Fig. 3.
figure 3

SMO ROC curve in WBC Dataset.

Full size image

In terms of the WBC dataset, our proposed method is compared with two studies [6, 10]. Results shows that the performance of SMO classifier is better since our model employs pre-processing, and resampling approaches. Thus, utilizing pre-processing, and resampling techniques play an important role in increasing the SMO accuracy comparable to the other techniques in [6, 10]. Details are shown below in Table 7.

Table 7. Compression of accuracy measures for the WBC Dataset.
Full size table

6 Conclusion

Breast cancer is considered to be one of the significant causes of death in women. Early detection of breast cancer plays an essential role to save women’s life. Breast cancer detection can be done with the help of modern machine learning algorithms. In this paper, we focus on how to deal with imbalanced data that have missing values using resampling techniques to enhance the classification accuracy of detecting breast cancer. In our work, three classifiers algorithms J48, NB, and SMO applied on two different breast cancer datasets. Results show that using the resample filter in the preprocessing phase enhances the classifier’s performance. In the future, the same experiments will apply to different classifiers and different datasets.