Modified fuzzy rough set technique with stacked autoencoder model for magnetic resonance imaging based breast cancer detection

ABSTRACT


INTRODUCTION
In the current scenario, breast cancer is the common cancer type in the rural and urban areas, where women between the age group of thirty-fifty years are at a higher risk of breast cancer [1], [2].It is the second most cause of cancer deaths in women after lung cancer [3].Hence, the death rate of women due to breast cancer is 1 in 37 subjects, which is around 2.7%.Therefore, the proper treatment and early diagnosis of breast cancer are essential for decreasing the death rates and preventing the disease progression [4]- [6].In recent decades, magnetic resonance imaging (MRI) images are highly utilized for diagnosing breast cancer to decrease unnecessary biopsies [7], [8].Additionally, the MRI images are a highly recommended test to monitor and detect the breast cancer lesion and to interpret the lesioned region, because it has better soft tissue imaging [9].Additionally, an experienced physician is needed to process the MRI images, which is a time-consuming mechanism [10], [11].For overcoming the above-stated issue, several automated models are implemented by the researchers [12], [13].Singh et al. [14] introduced a novel two-stage model for tumor classification.The integration of adversarial network and convolutional neural network (CNN) requires a large amount of medical data for training the developed model, which was extremely expensive.Ibraheem et al. [15] combined two dimensional median filter and discrete wavelet transform for improving the quality of breast images and extracting the features.The extracted features were given to the support vector machine (SVM) for tumor and healthy region classification.The SVM does not work well, when the target-classes were overlapping and the collected data was noisier.Khan et al. [16] introduced a deep learning framework based on the concept of transfer learning for breast cancer detection.In the presented deep learning system, three pre-trained models like residual network (ResNet), GoogLeNet and visual geometry group network (VGGNet) were used to extract features from the breast cytology images.The extracted deep learning features were fed to the fully connected layer of the transfer-learning model for malignant and benign classification.The developed transfer-learning model needs expensive graphics processing unit systems that increase computational cost.Ragab et al. [17] has integrated ResNet, AlexNet, and GoogLeNet models for extracting deep features from the breast mammogram images.The extracted features were given to the SVM classifier for tumor and non-tumor region classification.However, the SVM classifier was suitable for binary class classification, where it was inappropriate for multiclass classification.On the other hand, Alanazi et al. [18] has presented a CNN model for boosting the automatic detection of cancer regions by utilizing histopathology images, where it was computationally expensive.Fang et al. [19] firstly applied median filtering technique for enhancing the quality of mammogram images.Then, the whale optimization algorithm was combined with the multilayer perceptron algorithm for classifying the breast images as healthy or cancerous.The evaluation outcomes demonstrated that the presented model obtained higher accuracy than the existing models.The multilayer perceptron algorithm was sensitive to feature scaling and needs more hyper-parameters tuning such as hidden layers and neurons.Gravina et al. [20] developed a CNN model based on the intrinsic deforming autoencoders for automatic breast lesion malignancy recognition.The CNN was computationally costly, where it requires an enormous amount of data in order to obtain better classification results.Chouhan et al. [21] developed a deep highway network to extract dynamic features from the digital breast mammogram images.Further, the extracted features were given to the SVM and emotional learning inspired ensemble model for benign and malignant classification.As specified earlier, the developed SVM model supports only binary class classification.Khamparia et al. [22] has implemented a hybrid transfer-learning model that combines ImageNet and modified VGGNets for superior breast cancer recognition.The presented hybrid transfer-learning model was a superior tool for clinicians in order to diminish the false positive and false negative rates of breast cancer recognition, but it was computationally complex.In addition, Yurttakal et al. [23] implemented a time saving deep CNN model for classifying the breast lesions as benign or malignant tumors.The presented time saving deep CNN model obtained promising results in the breast cancer classification by means of specificity, accuracy, and sensitivity.In addition, Hizukuri et al. [24] developed a deep CNN model with Bayesian optimization for effective breast cancer classification.As presented in the resulting section, the deep CNN model obtained higher classification performance and it would be useful in early diagnoses of breast masses.However, the vanishing gradients was a major problem in the hybrid transfer-learning model and deep CNN model.To highlight the above-stated issues and to enhance breast-cancer detection, a novel deep learning system is implemented in this work.The primary aim of this article is to categorize the malignant and benign breast lesions with limited system complexity and computational time.The contributions are listed below: − After acquiring the breast images from breast imaging-reporting and data system (BI-RADS) MRI  This article is arranged in the following manner; the proposed methodology explanations are described in section 2. The simulation outcomes and its comparison are demonstrated in section 3. Finally, the conclusion of this study is depicted in section 4.

METHOD
In the context of breast cancer detection, the presented deep-learning framework comprises six phases.These six phases include image acquisition: BI-RADS MRI and DCE-MRI datasets, image denoising: region growing and adaptive histogram equalization, segmentation: Otsu thresholding with morphological transform, feature extraction: steerable pyramid transform and local ternary pattern descriptor, feature optimization: modified fuzzy rough set, and classification: stacked autoencoder.The workflow of the developed deep learning framework is represented in Figure 1.

Image acquisition and denoising
In this research work, the developed deep learning framework's effectiveness is validated using the BI-RADS MRI and DCE-MRI datasets.The BI-RADS MRI dataset consists of 200 MRI breast images in that 98 breast images are benign and 102 breast images are malignant.In the BI-RADS MRI dataset, the benign breast images are 17.63±5.79mm in size and the malignant breast images are 29.80±9.88mm in size.In addition, the subjects with granulomatous mastitis and infection are excluded from the research.Table 1 states the data statistics of the BI-RADS MRI dataset.On the other hand, the DCE-MRI dataset comprises 56 MRI examinations of 56 patients in which 30 are malignant masses and 26 are benign masses.The sample-acquired images are denoted in Figure 2.  histogram values to distinguish the images into many sections and then utilizes these sections for redistributing the lightness of the breast MRI images.Hence, the AHE technique is appropriate to enhance the image edges and to improve the local contrast of the collected breast MRI images [25].Additionally, the region growing technique completely relies on the neighborhood image pixel assumption [26].The region growing techniques compare one pixel with the neighbourhood pixels.If the similarity criterion is satisfied, the pixels belong to the clusters.The output image of the AHE and region-growing techniques are graphically depicted in Figures 3 and 4.

Tumor segmentation
After denoising the breast images, the tumor segmentation is accomplished by utilizing Otsu thresholding technique.The Otsu thresholding is an effective and simple segmentation technique, where it uses maximum class variance values.Related to existing image segmentation techniques, the Otsu thresholding technique includes advantages like need only smaller storage space, faster processing speed and ease in implementation.The pixel intensity level of the denoised image  is initially determined by utilizing (1).
where,    indicates the pixel intensity value that corresponds to the intensity levels from  until , and    represents distribution probability value of the denoised image.Additionally,  indicates image components (grayscale or RGB) and  specifies the number of pixel values in the denoised breast images [27].Next, the histogram values in the probability distribution are normalized using (2).(3) states average rate for the class variants one and two.Then, the objective function is calculated utilizing (5).
where, ℎ = ℎ 1 , ℎ 2 , … … .ℎ −1 represents a vector, which contains multiple thresholds.The Otsu thresholding between the class variance function is maximized to achieve the optimum threshold level of breast image for better tumor segmentation by increasing the objective function.In addition, the morphological dilation operator is employed on the output images of the Otsu thresholding technique that utilizes a structural element for expanding and probing the shapes in the output images of the Otsu thresholding technique.The output images of Otsu thresholding with morphological transform are graphically represented in Figure 5.

Feature extraction with optimization
After tumor segmentation, the feature extraction is performed by utilizing steerable pyramid transform (SPT) and local ternary pattern (LTP) for extracting the features from the segmented tumor regions.The SPT is a linear multi-orientation and multi-scale image decomposition method, where the major portions of the linear transforms are sub-band transforms.Initially, the SPT is an effective image decomposition method that partitions the segmented tumor images into numerous sub-bands using orientation and scale, which is calculated using decimation and convolution operations.The sub-bands of the SPT are rotation invariant and translation that reduce the concern of orthogonal-separable-wavelet-decomposition [28].In the SPT method, the segmented image is partitioned into low and high pass sub-bands utilizing the filters 0 and 0.Further, the low pass sub-bands are decomposed into four oriented band-pass sub-bands utilizing low pass filter 1 and band pass filters 0, 1, 2, and 3.Lastly, a robust image representation is generated with high orientation and scale by increasing the number of pyramid levels and number of image orientations.
Additionally, the LTP is a three-value texture descriptor for extracting the textual feature vectors from the segmented images.The LTP labels the image pixels with a threshold value by multiplying and adding the centre neighborhood image pixels   to generate the new labels.After defining the threshold value , the pixel values within the range of − to + are considered to assign the value of zero to the image pixels [29].The value 1 is assigned to the image pixels, if the value is higher than the threshold value, and the value -1 is assigned to the image pixels if the value is lower than the center pixel value.The mathematical expression of the LTP operator is represented in (6). where

Classification
After choosing the active features, the stacked autoencoder is applied for classification [30].The stacked autoencoder classification technique comprises multi-layer autoencoders that obtain higher-level representation of the original feature vectors by reconstructing input and its structure.In the input layer, the original information is encoded for obtaining the higher-level feature vectors of the middle-hidden layer and then the input information is reconstructed by decoding the information.By minimizing the reconstruction error, the stacked autoencoder networks are trained.The original training data is considered as  and the hidden layer is mathematically expressed in (7).
where  = ℎ (. ) represents activation function.Further, the output  is obtained by decoding the original information, which is mathematically represented in (8).Then, the objective is minimized for training the autoencoder that is defined in (9).
The stacked autoencoder is trained on the basis of layer-by-layer greedy method.Particularly, the feature vector of the upper hidden layer is used as the input of the succeeding layers, which is named as pretraining.Further, the weights of the pre-trained network are connected and then the weights of the final network are obtained by fine-tuning.The assumed parameters are: maximum iterations: softmax learning is 100, sparsity proportion is 0.15, maximum iterations: SAE learning is 100, L2 weight regularization is 0.004, sparsity regularization is 4, and a number of hidden layers is 100.

RESULTS AND DISCUSSION
In the automated breast cancer detection, the developed rough set based stacked autoencoder model's efficacy is simulated by MATLAB 2020.The rough set based stacked autoencoder model's efficacy is analyzed in terms of PPV, specificity, accuracy, NPV and sensitivity on the BI-RADS MRI and DCE-MRI datasets.However, the NPV and PPV are defined as the proportions of the negative and positive results in the statistic and diagnostic tests, which are true negative and true positive results.The formulae of NPV and PPV are depicted in equations ( 10) and (11).
The sensitivity is a test that determines the patient cases precisely and the specificity is a test that precisely identifies the healthy cases.In addition, accuracy is a vital evaluation metric in breast cancer detection that indicates how closer the obtained results are to the actual results.The mathematical formulae of sensitivity, specificity and accuracy are stated in ( 12), (13), and (14).Where, FP, TP, FN, and TN indicate false positive, true positive, false negative and true negative values.

Quantitative analysis
In this scenario, the rough set based stacked autoencoder model's efficacy is analysed on the BI-RADS MRI dataset that comprises 200 MRI breast images in which 20% breast images are utilized for model testing.Additionally, the proposed rough set based stacked autoencoder model is evaluated by a five-fold cross validation technique that diminishes the computational time, and the variance of the estimated parameters against the dataset.In Table 2, the experimental analysis is carried-out utilizing different classifiers such as stacked autoencoder, naïve Bayes, random forest (RF), decision tree (DT), and k-nearest neighbor (KNN) along with and without feature optimization in that the combination: a modified fuzzy rough set technique with stacked autoencoder obtained maximum performance with sensitivity of 98.68%, PPV of 98.11%, classification accuracy of 99%, specificity of 99%, and NPV of 97.56% on the BI-RADS MRI dataset.The experimental outcome by changing the classifiers with and without feature optimization is represented in Figures 7 and 8. Table 3 reveals that the modified fuzzy rough set technique with stacked autoencoder achieved maximum classification performance in the breast cancer detection on the BI-RADS MRI dataset related to other optimizers like artificial bee colony (ABC), particle swarm optimizer (PSO), and grey wolf optimizer (GWO).In this research manuscript, the modified fuzzy rough set technique is significant for visualization and feature optimization of the BI-RADS MRI dataset, where it effectively deals with the hybrid decision systems.The experimental results by varying the feature optimizers are graphically stated in Figure 9.
Similar to the BI-RADS MRI dataset, the proposed rough set based stacked autoencoder has obtained higher classification accuracy of 99.22%, sensitivity of 98.80%, specificity of 98.91%, PPV of 98.92%, and NPV of 98.90% on the DCE-MRI dataset.The achieved experimental results are maximum related to other classifiers and optimizers, as specified in Tables 4 and 5.

Comparative analysis
The comparative results of the prior models and the rough set based stacked autoencoder model is represented in Table 6.Yurttakal et al. [23] presented a deep CNN model for classifying the breast MRI lesions as malignant and benign.The simulation outcomes demonstrated that the deep CNN model achieved 98.33% of classification accuracy and 96.88% of specificity on the BI-RADS MRI dataset.In addition, Hizukuri et al. [24] integrated a deep CNN model with Bayesian optimization for effective breast cancer classification.As depicted in the resulting section, the developed model obtained 92.90% of accuracy and 92.30% of specificity on the DCE-MRI dataset.Compared to the existing deep CNN models, the rough set based stacked autoencoder model obtained superior performance in breast lesion detection by means of different evaluation measures.

Discussion
The segmentation, feature optimization, and classification are the vital parts of this research for precise breast cancer detection with minimal computational time.The Otsu thresholding with morphological transform is simple and speed in finding the optimal threshold value for separating foreground and background regions.After feature extraction, the incorporation of the modified fuzzy rough set technique in the proposed system diminishes the computational time and complexity of the stacked autoencoder by selecting active features.The computational complexity of the rough set based stacked autoencoder model is linear (), where  indicates the order of magnitude and  states input size.The proposed model consumes 34 and 22 seconds to train and test the BI-RADS MRI and DCE-MRI datasets, which are minimum related to other comparative classification models.Additionally, the stacked autoencoder easily and effectively learns non-linear transformation with multiple layers and activation functions for better disease classification.The computational complexity and training time are the major issues addressed in the related work section that are effectively overcome by the rough set based stacked autoencoder model.

CONCLUSION
In this manuscript, a new rough set based stacked autoencoder model is implemented for effective breast cancer detection.The aim of this research is to develop a feature optimizer and an effective deep learning classifier for effective classification of breast cancer.Therefore, the most pre-dominant Int J Elec & Comp Eng ISSN: 2088-8708  Modified fuzzy rough set technique with stacked autoencoder model for … (Sachin Kumar Mamdy) 295

Figure 1 .
Figure 1.Workflow of the developed deep learning framework

Figure 5 .
Figure 5. Output images of Otsu thresholding with morphological transform 299 the modified fuzzy rough set technique is proposed that integrates the fuzzy equivalence and the membership function of fuzzy c means clustering technique for feature optimization.The modified fuzzy rough set technique knows the dataset and selects the highly correlated feature vectors of 537 for disease classification.The flowchart of the modified fuzzy rough set technique is represented in Figure 6.

Figure 6 .
Figure 6.Flowchart of the modified fuzzy rough set technique

ISSN: 2088- 8708 Figure 7 .Figure 8 .
Figure 7.Comparison result by changing the classifiers without feature optimization on the BI-RADS MRI dataset Int J Elec & Comp Eng ISSN: 2088-8708  Modified fuzzy rough set technique with stacked autoencoder model for … (Sachin Kumar Mamdy) 303 discriminative feature vectors are selected utilizing the modified fuzzy rough set technique.Further, the selected features are given as the input to the stacked autoencoder for classifying both the malignant and benign breast lesions.The proposed rough set based stacked autoencoder model delivers superior performance in the breast cancer recognition in terms of classification accuracy, PPV, specificity, NPV, and sensitivity.In the experimental segment, the proposed rough set based stacked autoencoder model obtained classification accuracy of 99% and 99.22% on the BI-RADS MRI and DCE-MRI datasets.The obtained experimental outcomes are superior to the conventional classifiers and optimizers.Breast cancer detection by the proposed rough set based stacked autoencoder model can assist doctors and pathologists in the classification of abnormalities with maximum accuracy in minimal computational time.In future, a new ensemble based deep learning model can be included in the proposed system to further detect the sub-stages of breast cancer.

Table 1 .
Data statistics of the BI-RADS MRI dataset ,  2 represents variants between the classes  (benign and malignant classes),  1  and  2  denotes class variants one and two, and  0  and  1  ISSN: 2088-8708 Int J Elec & Comp Eng, Vol.14, No. 1, February 2024: 294-304 ,  indicates user-specified threshold,   denotes neighborhood pixel value and   represents a central pixel value.The hybrid feature extraction (LTP and SPT) extracts 1536 features from the segmented images.
The dimensions of the extracted feature vectors are decreased by implementing the modified fuzzy rough set technique.Generally, the fuzzy rough set utilizes two approximations such as lower and upper limits for feature optimization that ranges between [0, 1].The conventional fuzzy rough set feature optimization is employed on nominal, valued, continuous and nominal data, where it significantly handles the data noise.Several reformulations are carried out in a fuzzy rough set to speed up the calculations.In this manuscript, ISSN: 2088-8708  Modified fuzzy rough set technique with stacked autoencoder model for … (Sachin Kumar Mamdy)

Table 2 .
Experimental result by changing the classifiers on the BI-RADS MRI dataset

Table 3 .
Experimental result by changing the feature optimizers on the BI-RADS MRI dataset

Table 4 .
Experimental result by changing the classifiers on the DCE-MRI dataset

Table 5 .
Experimental result by changing the feature optimizers on the DCE-MRI dataset

Table 6 .
Comparative evaluation between the existing and the proposed rough set based stacked autoencoder model