Mask Branch Network: Weakly Supervised Branch Network with a Template Mask for Classifying Masses in 3D Automated Breast Ultrasound

Abstract

Automated breast ultrasound (ABUS) is being rapidly utilized for screening and diagnosing breast cancer. Breast masses, including cancers shown in ABUS scans, often appear as irregular hypoechoic areas that are hard to distinguish from background shadings. We propose a novel branch network architecture incorporating segmentation information of masses in the training process. The branch network is integrated into neural network, providing the spatial attention effect. The branch network boosts the performance of existing classifiers, helping to learn meaningful features around the target breast mass. For the segmentation information, we leverage the existing radiology reports without additional labeling efforts. The reports, which is generated in medical image reading process, should include the characteristics of breast masses, such as shape and orientation, and a template mask can be created in a rule-based manner. Experimental results show that the proposed branch network with a template mask significantly improves the performance of existing classifiers. We also provide qualitative interpretation of the proposed method by visualizing the attention effect on target objects.

1. Introduction

Breast cancer is the second leading cause of cancer death in women worldwide [1,2]. Many studies have revealed that breast screening can discover cancer in its early stages, reducing mortality [3,4,5,6,7]. Mammography has been the most widely used examination for breast screening, but it often misses cancer hidden in dense breast tissue that contains much fibrous or glandular tissue rather than fat [8]. Hand-held ultrasound (HHUS) imaging is a popular alternative covering dense breast tissue, but it suffers from low reproducibility depending on its operator. Recently, automated breast ultrasound (ABUS) has been introduced and is receiving favorable reviews [9]. While an HHUS makes partial two-dimensional breast scans that need to be simultaneously examined by an operator on site, an ABUS produces whole three-dimensional (3-D) breast scans with a dedicated probe that can be asynchronously examined later. This has led to the active development of computer-aided diagnosis (CAD) systems for ABUS, which can assist interpreting physicians by reducing their workload and enhancing cancer detection performance [10,11,12].

The main task of CAD systems is to detect breast masses that can possibly grow into cancer. Breast masses on ABUS scans are usually shown as hypoechoic areas that can also appear due to a variety of causes, such as fat, shadow, and anechoic tissue; thus, many CAD systems erroneously detect these areas as suspicious breast masses, resulting in false positives. Even radiology professionals have difficulty distinguishing hypoechoic areas of breast masses from those of other causes and have been found to detect a number of false positives on breast ultrasound scans [13,14]. Several studies [15,16,17,18] have proposed image classifiers based on conventional image processing techniques but have not been reported to significantly reduce false positives.

To develop competitive classifiers on ABUS, recent studies have employed convolutional neural networks (CNNs) that have shown tremendous success in image classification tasks. Chiang et al. [19] designed a two-stage breast mass detector on ABUS scans with a modified CNN architecture. A fast sliding window first searches for volumes of interest, which is then fed into a simple three-dimensional extension of a 2-D CNN to compute probabilities of tumor candidates. Moon et el. [20] incorporated focal loss and ensemble learning into their 3-D CNN model to resolve a data imbalance issue. Although those methods have been shown to sensitively detect breast masses, they yield an average of more than four false positives per scan, which means room for improvements still remains. Simple modifications of the existing CNNs may not be sufficient to analyze the three-dimensional context of breast masses on ABUS scans.

Training a CNN requires many sample images that preferably contain clear features of target objects that can be distinguished from their backgrounds. Compared to ordinary objects (e.g., cats, boats) shown in common training samples, breast masses shown in ABUS scans may not present clear boundaries. As seen in the first row of Figure 1, the breast masses (indicated by the arrows) often appear as irregularly shaded blobs, and each blob itself is hardly distinguishable from its background shadings. Moreover, the inherent low image resolution and high noise level of ABUS scans make it more difficult to train CNNs for indistinct masses.

Considering these characteristics, we empirically found that training with a segmentation mask on a mass helps improve the classification performance, probably because the low-level features on the blobs are activated on the masses rather than on the backgrounds. The second row of Figure 1 shows class activation maps (CAMs) for mass classification using the existing DenseNet [21]. The activated network weights are widely distributed across the background blobs and are not concentrated on the masses. In contrast, our modified classifier with segmentation masks intensively activates the weights on the mass blobs, as illustrated in the third row. To integrate the mask information, we attached an additional network (called the mask branch network) to the middle of a classification network. Since our branch network is independently applicable, the proposed method can be modified from any existing state-of-the-art CNN model.

A practical issue of the proposed method lies in the difficulty of generating segmentation masks Aside from the high cost of professional tasks within a three-dimensional space, the exact boundary is not clear even for medical experts. We propose to employ a template mask with predefined shapes instead (e.g., circle, oval) with minimal parameters such as diameter and direction. This information is typically recorded in radiology reports; thus, additional labeling is rarely required.

The contributions of the proposed method can be summarized as follows:

• Mask Branch Network: We propose a branch network incorporating mask information that can be attached to the middle of existing networks yielding intermediate features of input scans.
• Template Mask: We use the characteristics of breast mass recorded in radiology reports to generate a simplified segmentation mask, requiring no additional labor for segmentation.
• Performance evaluation: Extensive experiments exhibit that the proposed method contributes to improve the model’s performance. Especially, We provide considerable test results with different architectures of the branch network as well as with variations of the template mask in the process of obtaining an optimal model.

2. Materials and Methods

The overall architecture of the proposed method is illustrated in Figure 2. Our network architecture consists of two parallel networks: a main network and a mask branch network. The main network works as a backbone CNN classifier. Any high-performance CNN classifier can be a main network, for which we employ a 3-D image classifier based on the DenseNet [21] architecture. In the middle of the main network, the mask branch network bifurcates off to compute the mask loss by comparing it with the template mask generated from a radiology report. The main loss and the mask loss from each branch are integrated into the total loss to be optimized in the training process.

2.1. Main Network

The main network implements a binary classifier determining the existence of suspicious breast lesions. It is composed of four dense blocks with transition layers based on DenseNet-BC-121 [21]. A transition layer is placed between two consecutive dense blocks to perform downsampling, batch normalization, a $1\times 1\times 1$ convolution, and average pooling. The global average pooling (GAP) layer is connected at the end of the last dense block, followed by the fully connected (FC) layer and the softmax activation layer sequentially. The main loss can be defined as a cross entropy loss of input samples, shown as follows:

$$L_{main}=\frac{1}{N}\sum _i^N[(1-y_i)log(1-p_i)+y_{i}log{p}_i],$$

(1)

where i means the index of an input sample, N is the number of samples, $y_i\in \{0,1\}$ is the ground-truth label, and $p_i$ is the probability from the main network estimating whether the input sample contains a mass.

2.2. Mask Branch Network

The concept of branch architecture [22,23] is to train multiple tasks that take advantage of the interaction between different tasks. The mask branch aims to integrate spatial information into the main network, which helps the main network extract meaningful features from the area around the mass.

The mask branch is designed to start at the $1\times 1\times 1$ convolutional layer of the second transition layer of the main branch and to generate a branch output, which is a voxel-level probability map that estimates the presence of target lesions. We may assume that this branch output simulates a segmentation mask.

We denote $\mathbf{F}\in {\mathbb{R}}^{C\times H\times W\times D}$ as a midlayer CNN feature extracted from the second transition layer, where $C,H,W$ and D are the number of channels, height, width and depth of the feature map, respectively.

The branch output B can be formulated as:

$$\mathbf{B}=softmax\left(f\right(\sigma \left(\mathbf{F}\right),w\left)\right),$$

(2)

where $f(·,·)$ denotes a 3D convolution function, $\sigma$ represents the activation function, $\mathbf{B}\in {\mathbb{R}}^{H\times W\times D\times 2}$ means the generated branch output map, and $w\in {\mathbb{R}}^{2\times C\times 1\times 1\times 1}$ indicates convolutional parameters. As a result, branch output is a voxel-level binary probability map of the target lesion and background.

Comparison between branch output and the mask yields the mask loss, a pixelwise cross-entropy loss, by which the model learns pixelwise class probabilities. With the mask that contains spatial information of the target lesion, the mask loss is calculated pixelwise between the branch output and the mask. The definition of mask loss is based on the equation below:

$$L_{mask}=\frac{1}{NM}\sum _j^N\sum _i^M[(1-z_{i,j})log(1-q_{i,j})+z_{i,j}log{q}_{i,j}],$$

(3)

where j is the index of a sample of the input batch, i is the index of a voxel in the $H\times W\times D$ space, $M=HWD$ is the total number of voxels in the branch output, $z_{i,j}\in \{0,1\}$ is a pixel value of the mask $\mathbf{T}\in {\mathbb{R}}^{H\times W\times D}$, which is proposed in Section 2.4, and $q_{i,j}$ is the probability from the mask branch.

The total loss of the whole model, subject to optimization, is a combination of main loss and mask loss. To combine these losses efficiently, the uncertainty loss weight method [24] commonly used for multitask loss is adapted for combining our losses and expressed as follows:

$$L_{total}=\frac{1}{{\sigma }_1^2}{L}_{main}+\frac{1}{{\sigma }_2^2}{L}_{mask}+log{\sigma }_1+log{\sigma }_2,$$

(4)

where ${\sigma }_1$ and ${\sigma }_2$ are trainable variables that learn the relative weights of $L_{main}$ and $L_{mask}$ adaptively and regulate the balance of the losses. When training model, $L_{total}$ let scale value of ${\sigma }_1$ and ${\sigma }_2$ regulate the contribution of $L_{main}$ and $L_{mask}$, preventing ${\sigma }_1$ and ${\sigma }_2$ being too large.

Note that the mask branch network is used only for the training process to focus on the region of interest and not for the testing process. Thus, the proposed method does not increase computational complexity in the inference phase.

2.3. Variation of Network Designs

We designed a mask branch network with four types from MBN-V1 to MBN-V4, as shown in Figure 3. We explored the architecture of the mask branch network with several factors: the number of branches, starting positions of branches, and how to combine branch outputs in the case of multiple branches. Although the principle of operation is the same as that explained by MBN-V1 above, the suitable structure of the mask branch network can be diverse for each backbone network and task. We design single or multiple subnetworks parallel to the main network. MBN-V1 and MBN-V2 have a single branch, generating only one output map.

MBN-V1 is attached to the 2nd transition layer, extracting feature map ${\mathbf{F}}_{\mathbf{1}}$ and resulting in ${\mathbf{B}}_{\mathbf{1}}$ (Figure 3a). The mask loss $L_{mask}$ is used to reduce the difference between ${\mathbf{B}}_{\mathbf{1}}$ and ${\mathbf{T}}_{\mathbf{1}}$. MBN-V2 is integrated into the backbone on the third transition layer, as shown in Figure 3b. In contrast to MBN-V1 and MBN-V2, the other versions of the mask branch network consist of two branches that were used for MBN-V1 and MBN-V2. In MBN-V3, the two branches are separated, resulting in ${\mathbf{B}}_{\mathbf{1}}$ and ${\mathbf{B}}_{\mathbf{2}}$, which have different shapes, as shown in Figure 3c. The branch outputs are matched to ${\mathbf{T}}_{\mathbf{1}}$ and ${\mathbf{T}}_{\mathbf{2}}$, calculating two corresponding mask losses. Finally, the losses are combined with the uncertainty loss weight method, and the branches are optimized by each loss. Conversely, MBN-V4 integrated the output maps to one single array by a summation operation (Figure 3d). In MBN-V4, after ${\mathbf{B}}_{\mathbf{1}}$ is downsampled to the same shape as ${\mathbf{B}}_{\mathbf{2}}$ through 3D average pooling, ${\mathbf{B}}_{\mathbf{1}}$ and ${\mathbf{B}}_{\mathbf{2}}$ are integrated by elementwise summation.

We show the evaluation results of each case implemented in the experiments section.

2.4. Template Mask

The shading blobs in the background ABUS image are often indistinguishable from the shading of the mass. For the segmentation mask, a pixelwise segmentation map fully annotated by clinical experts can be an ideal mask; however, manually tracking the boundaries of a three-dimensional mass requires considerable time and effort from the expert. To address this annotation problem, we propose to utilize the information that already exists in the radiology report instead of manually labeling the ground truth.

2.4.1. Criteria on Radiology Reports

The location, size and category of suspicious lesions are usually written in radiology reports during diagnostics. By annotating the three axes of the lesion, its location and size are recorded in the form of center coordinates and diameters. For categories, the Breast Imaging Reporting and Data System (BI-RADS) [25] category is the most widely used standard for determining the malignancy of breast lesions. As factors for BIRADS, some visual characteristics are recorded in radiology reports during the diagnosis of breast images; ‘shape’, ‘orientation’, ‘margins’, ‘posterior acoustic feature’, etc. [25] Among those properties, we employ orientation features, indicating that the direction of the long axis relative to the breast skin is parallel or not because it reflects the approximate shape of the lesion with two kinds of categories: parallel or not parallel. By utilizing these properties, we can automatically create a simplified label mask, named the template mask, which contains spatial information and guides the classifier where to focus. Due to the limited information available, this simple mask cannot be regarded as an exact segmentation label map obtained by the annotation process.

2.4.2. How to Create a Template Mask

Similar to the general segmentation map, the template mask is a binary label map that consists of positive and negative areas. The positive area is created by the rule utilizing the position, diameter, and orientation of breast lesions. The shape of the template mask is one of two types: sphere or ellipsoid. If the lesion has a ‘parallel’ property of orientation, the diameter perpendicular to the skin is halved, changing the shape of the positive area from a sphere to an ellipsoid. With the center coordinate $(x_t,y_t,z_t)$, diameter $d_t$ and orientation of the lesion, we can define a template mask $\mathcal{T}(x,y,z)$ as a simplified segmentation label.

$$\mathcal{T}(x,y,z)=\left\{\begin{array}{cc}1\hfill & \mathrm{if}\frac{{(x-x_t)}^2}{d_t}+\frac{{(y-y_t)}^2}{\alpha d_t}+\frac{{(z-z_t)}^2}{d_t}\le 1,\hfill \\ 0\hfill & \mathrm{else}\hfill \end{array}\right.$$

(5)

where $\alpha$ is 0.5 if the orientation is parallel; otherwise, $\alpha =1$.

An example of a template mask is shown in Figure 4 with a sampled breast cancer. Figure 4a–c show transverse, coronal, and sagittal views of the center of a breast lesion. With lesion diameters of 10 mm, 18 mm, and 15 mm on each axis, the average diameter was 14.3 mm. The centroid of the template mask is located in the same position as the center of the lesion. Since the lesion has a ‘parallel’ feature of orientation, the positive area shrinks in the skin-perpendicular direction, obtaining the shape of an ellipsoid whose semidiameters are one of 7.15 mm and two of 14.3 mm, as shown in Figure 4d–f.

3. Results

To the best of our knowledge, no public dataset is available for ABUS research; thus, we built our own dataset in a tertiary hospital. ABUS images used in this study were acquired from ABUS systems (Invenia ABUS, Automated Breast Ultrasound System; GE Healthcare, Sunnyvale, CA, USA). For each breast, three volumes were obtained: the central volume, the lateral volume, and the medial volume. The institutional review board approved this study and waived informed consent, considering the retrospective study design and the use of anonymized patient data.

3.1. Implementation Details

A total of 363 patients who underwent ABUS from May 2017 to October 2019 were included. ABUS images of 286 patients presented 434 mass lesions categorized as $C2$ by BI-RADS, while other 77 images showed no mass lesions. We randomly divided the mass lesions: 304 lesions in the training set, 50 lesions in the validation set, and 80 lesions in the test set. The dataset also includes $3907$ nonmass lesions that were randomly cropped from the 77 normal patients’ images and randomly divided into $3777$ lesions for the training set and 50 and 80 lesions for the validation and test sets, respectively.

All center coordinates and diameters of masses were obtained from radiology reports annotated by experienced radiologists. We rescaled the original volume images to have voxel sizes of $0.3$ mm to $2.0$ mm depending on the mass size, and then we cropped the network input volumes at the center coordinate of the mass with a size of $48\times 32\times 48$ voxels. Additionally, the 3-dimensional patches used for training are rotated three times with rotation angles of ${90}^{\circ }$, ${180}^{\circ }$, and ${270}^{\circ }$ for augmentation.

The weights of the convolutional layers were initialized by the Xavier uniform initializer [26]. We performed batch normalization (BN) [27] immediately after each convolution layer, and ReLU activation was followed by BN. The models we used were trained with a batch size of 48 with 2 mini batches of the same size, and the losses were optimized by the RMSprop optimizer [28]. All networks in this study were implemented with the TensorFlow 1.12 library and were trained on Nvidia 2080-Ti on an Ubuntu 18.04 system.

3.2. Evaluation Metrics

The classification performance was evaluated by measuring the sensitivity ($Se$), specificity ($Sp$), and area under the curve ($AUC$) of receiver operating characteristics ($ROC$). The sensitivity and specificity are the percentage of positive and negative results that are correctly identified. The metrics are defined as:

$$Sensitivity\left(Se\right)=\frac{TP}{TP+FN}$$

(6)

$$Specificity\left(Sp\right)=\frac{TN}{TN+FP}$$

(7)

where $TP$ and $TN$ are true positive and true negative and likewise FP and FN are false negative and false positive. The AUC was calculated by using receiver operating characteristic (ROC) analysis on the test set.

3.3. Quantitative Results

3.3.1. Variation of MBN Architectures

To evaluate the effect of the proposed mask branch network (MBN), we tested the main network as a baseline and compared the results from the main network with MBN. We employed two networks as the main network: DenseNet-BC [21] and ResNetV2-101 [29]. As shown in

Table 1. Performance comparison with the mask branch network(MBN) on DenseNet-BC and ResNetV2-101 on mass classifier.

	Sensitivity	Specificity	AUC
DenseNet	56.25%	86.25%	0.8553
DenseNet + MBN-V1	75.00%	92.50%	0.9252
DenseNet + MBN-V2	87.75%	93.75%	0.9491
DenseNet + MBN-V3	82.50%	92.50%	0.9472
DenseNet + MBN-V4	77.50%	92.50%	0.9483
ResNet	65.00%	70.00%	0.7675
ResNet + MBN-V1	68.75%	77.50%	0.8270
ResNet + MBN-V2	71.25%	86.25%	0.8795
ResNet + MBN-V3	70.00%	91.25%	0.9261
ResNet + MBN-V4	67.50%	90.00%	0.8797

, the proposed networks with MBN outperform the baselines in every performance measure in both kinds of main networks. While all versions of our method achieve higher performance overall compared with the baseline, the version of best performance is different depending on the baseline. In the case of DenseNet, MBN-V2 achieved the best performance for all metrics. In the lower part of Table 1, the result of our method with ResNet is shown for performance comparison.

Similar to cases with DenseNet, all of the proposed methods outperform the plain ResNet model. Unlike experiments based on DenseNet, the mask branch network achieved the best performance by using MBN-V2 for sensitivity and MBN-V3 for specificity and AUC. Figure 5 compares ROC curves for the mass classifier with the mask branch network.

3.3.2. Comparison of Loss Weighting Strategies

We compared weighting loss methods with mask loss. Various methods have been used to efficiently combine multiple losses in multitask learning (MTL) [30]. The most straightforward method is uniform weighting: the losses are simply added together to produce a single scalar loss value. Dynamic optimization techniques are also important in MTL to optimize the set of possibly contrasting losses or gradients because conflicting gradient problems can degrade model performance. Kendall et al. [24] proposed an uncertainty-based weighting approach and applied it to a CNN. Ref. [31] adapted the regularization term in Kendall’s uncertainty-based method. A dynamic weight average (DWA) was proposed by [32]. We adapted these methods to the proposed method with MBN-V1 and exactly the same experimental setting and compared the performance. The results are shown in

Table 2. Experiments on ways to weigh the main loss and the mask loss. Metrics ranked with 1st place are in bold.

	Se	Sp	AUC
Equal Weighting	65.00%	82.50%	0.8720
Uncertainty Weighting	75.00%	92.50%	0.9252
Revised Uncertainty Weighting	71.25%	93.75%	0.9369
Dynamic Weight Average	70.00%	77.50%	0.8416

, which shows that the uncertainty weighting and revised uncertainty weighting achieve higher performance compared to other settings. The results do not show many differences between uncertainty weight and revised uncertainty weight, but their performance is considerably better than the others. Considering this result, we applied uncertainty weighting to our method for the final result. The left of Figure 6 shows ROC curves depending on ways to weigh multiple losses.

3.3.3. Characteristics Suitable for Template Mask

To evaluate the effect of mass features available in radiology reports, we compared the performance of template masks utilizing various feature sets.

Table 3. Comparison on set of characteristics utilized for creating Mask Label; position, diameter, orientation, posterior eco pattern. Metrics ranked with 1st place are in bold.

Template Label	Se	Sp	AUC
No mask	56.3%	86.3%	0.855
Loc.	67.5%	90.0%	0.908
Loc. + Size	72.5%	90.0%	0.921
Loc. + Size + Orien.	75.0%	92.5%	0.925

shows the cases using different features from breast lesions and their results. We adapted the cases to denset with MBN-V1 and uncertainty weighting. Even utilizing location only (with a fixed sphere of 20 mm diameter) helps to improve all performance metrics compared to the baseline (DenseNet without MBN; no mask) The size (diameter) also enhanced the metrics, and the combination of all features yielded the best performance, as seen in Table 3. In addition, The right graph of Figure 6 compares ROC curves for various template masks.

Figure 5. ROC curves of the mask branch network with DenseNet (Left) and ResNet (Right).

Figure 5. ROC curves of the mask branch network with DenseNet (Left) and ResNet (Right).

Figure 6. ROC curves of loss weight methods and template masks. (Left): Models using various ways to weigh multiple losses; (Right): Cases utilizing different characteristics for breast lesions.

Figure 6. ROC curves of loss weight methods and template masks. (Left): Models using various ways to weigh multiple losses; (Right): Cases utilizing different characteristics for breast lesions.

3.4. Qualitative Results

Figure 7 qualitatively visualizes the spatial attention effect of the proposed mask branch by plotting the grad-CAM (class activation map) visualization. The first row shows sample mass images with red rectangles indicating the masses. The CAM of DenseNet (second row) does not appropriately activate areas where the masses are present. In contrast, the proposed method (third row) shows relatively clear activation on the target areas; thus, the main network is expected to learn to exploit and aggregate features in the target area.

3.5. Cancer Classification

We also evaluated the performance of the cancer classification task, which determines whether the lesion is malignant, as shown in

Table 4. Cancer classification result on mask loss with DenseNet.

	Se	Sp	AUC
DenseNet	43.75%	56.25%	0.5187
DenseNet + MBN-V1	56.25%	60.41%	0.6402
DenseNet + MBN-V2	63.95%	61.57%	0.6802
DenseNet + MBN-V3	62.50%	64.58%	0.6050
DenseNet + MBN-V4	52.63%	68.42%	0.6638

. Due to the high similarity of benign and malignant lesions, the cancer classifier is a much more challenging task than the mass classifier. As seen in the first row of Table 4, without a mask branch network, the network is hardly trained with approximately 0.5 metrics. The performances of all of the cases with DenseNet are improved compared with the plain network. MBN-V2 looks the best-performing version in cancer classification, similar to mass classification in terms of sensitivity and AUC, although the highest specificity is from MBN-V4.

4. Discussion

The best-performing architecture of the mask branch network depends on the adopted tasks: mass classification and cancer classification. The mass classifier tended to result in higher performance when the mask branch network was attached to a deeper position, whereas the cancer classifier did not. The discrepancy depends on how similar the objectives of the main and branch networks are: The mask branch network has a similar function as the mass classifier in terms of detecting the existence of lesions, in which the mass classifier achieved greater improvement as more features are shared by attaching the mask branch network to the deeper position. However, the cancer classifier has to learn the malignancy of lesions, not the existence of lesions. Therefore, detailed features in layers at the bottom of the main network for cancer classification can be interrupted by the mask branch network. Due to the difference between the functions of networks, locating the attached position to the deeper position of the cancer classifier is better, reducing shared features.

A few kinds of information in radiology reports can be utilized for creating template masks. In the experiment section, we implemented and evaluated various types of template masks with orientation, position, and diameter. All cases with template masks showed better performance than the case without template masks.

In addition to input images, applying the proposed method requires detailed information on breast lesions with additional labor costs or costs. Hence, it is important to minimize additional information to increase efficiency. The objective of the template mask we intended is to cover the area where the breast mass exists. Therefore, no more information is needed if the mask is large enough to cover the target area. Thus, a template mask with minimal information, such as position and diameter, performed sufficiently.

5. Conclusions

This study introduced our novel branch network incorporating spatial information to improve the performance of a CNN for classifying masses on ABUS images. We utilized the characteristics of breast mass recorded in radiology reports to generate a simplified mask, a template mask that required no additional labor for segmentation. The proposed MBN can be attached to existing networks and has been shown to boost performance. We tested variations in the branch network and found the optimal architecture of the MBN. The proposed method was effective even in cancer classification, which is a more complicated task.