Biomedical image processing is a hot research topic which helps to majorly assist the disease diagnostic process. At the same time, breast cancer becomes the deadliest disease among women and can be detected by the use of different imaging techniques. Digital mammograms can be used for the earlier identification and diagnostic of breast cancer to minimize the death rate. But the proper identification of breast cancer has mainly relied on the mammography findings and results to increased false positives. For resolving the issues of false positives of breast cancer diagnosis, this paper presents an automated deep learning based breast cancer diagnosis (ADL-BCD) model using digital mammograms. The goal of the ADL-BCD technique is to properly detect the existence of breast lesions using digital mammograms. The proposed model involves Gaussian filter based pre-processing and Tsallis entropy based image segmentation. In addition, Deep Convolutional Neural Network based Residual Network (ResNet 34) is applied for feature extraction purposes. Specifically, a hyper parameter tuning process using chimp optimization algorithm (COA) is applied to tune the parameters involved in ResNet 34 model. The wavelet neural network (WNN) is used for the classification of digital mammograms for the detection of breast cancer. The ADL-BCD method is evaluated using a benchmark dataset and the results are analyzed under several performance measures. The simulation outcome indicated that the ADL-BCD model outperforms the state of art methods in terms of different measures.
Cancer is a disease which arises if anomalous cell grows in an uncontrolled way which disregards the average rules of cell partition that might create uncontrolled proliferation and growth of the anomalous cell. This could be fatal when the proliferation is permitted to spread and continue in this manner will result in metastases development [
Computer aided detection and diagnosis (CAD) systems are now being utilized for offering essential support in making decision procedures of radiotherapists. This system might considerably decrease the number of efforts required to the calculation of cancer in medical practice when minimizing the amount of false positives which result in discomforting and unnecessary biopsies. CAD system about mammography might tackle 2 distinct processes: diagnosis of detected lesions (CADx) and detection of suspected lesion in a mammogram (CADe), viz., classifications as malignant/benign [
This paper presents an automated deep learning based breast cancer diagnosis (ADL-BCD) model using digital mammograms. The proposed model involves Gaussian filter based pre-processing and Tsallis entropy based image segmentation. In addition, Deep Convolutional Neural Network based Residual Network (ResNet 34) is applied for feature extraction purposes. Specifically, a hyper parameter tuning process using chimp optimization algorithm (COA) is applied to tune the parameters involved in ResNet 34 model. The wavelet neural network (WNN) is used for the classification of digital mammograms for the detection of breast cancer. The ADL-BCD method is evaluated using a benchmark dataset and the results are analyzed under several performance measures.
Zhang et al. [
Yala et al. [
Shen et al. [
Kumar et al. [
Kaur et al. [
Masni et al. [
This study has introduced a new ADL-BCD technique to detect the occurrence of breast cancer using digital mammograms. The overall workflow of the ADL-BCD technique is showcased in
At the initial stage, the digital mammograms are preprocessed using GF technique to remove the unwanted noise. The Breast masses i.e., existing in the digital mammogram appear bright compared to the preprocessing, and background filters utilized must be capable of retaining its natural intensity features when eliminating the redundant noise portion. The presented scheme utilizes a 7 * 7 Gaussian filters i.e., a nonuniform lowpass filter to pre-process the digital mammograms in which the images are smoothened and the noises are detached in that way removing its intensity in homogeneity and preserve its grey level variation without the segmentation algorithm which might misinterpret for finding the actual breast masses [
Next, in the second stage, the pre-processed image is fed into the segmentation model to detect the lesion regions in the input mammograms. The entropy is associated with the chaos measure within a system. Initially, Shannon deliberated the entropy for measuring the ambiguity about the data content of the scheme [
According to Shannon concept, a non-extensive entropy theory was developed by Tsallis i.e., determined by:
The Tsallis entropy could be deliberated for finding an optimum threshold of an image. Assume that
In the multilevel thresholding procedure, it is essential for determining an optimum threshold value
During the third stage, the ResNet 34 model is used to effectually extract the features from the segmented images. Deep network is a multilayer neural network framework with multiple hidden layers. The learning of deep networks is generally performed hierarchically, starts from the low level to high level, using several layers of the network. DL depending on CNN was extensively utilized in many regions for solving distinct engineering challenges and displayed great performances in problem solutions [
CNN is a familiar model which permits the network for extracting local and global features from the data, improving the decision making process. At the convolution layer the value of all positions
Afterward addition the bias term
Every neuron will process a linear output. If the outcome of neuron is feed to other neurons, it finally makes other linear output [
This technique undergoes vanishing-gradient issues and containing huge computational difficulty Another non-linear activation function has TanH that was fundamentally a scaled version of the
That is to avoid the vanishing-gradient issue and their features. One of the famous non-linear operators was Rectified Linear Unit
The Leaky-ReLU rectifier features that are change of
As stated, the “in-depth” structure created an optimization complexity at the time of network training, that is gradient reducing problems and affect the network performances. In this work, they present residual learning to beat these problems and develop a DL framework to grade the fruits. In the residual network, the stacked layer performs a residual mapping by making shortcuts connections that carry out identity mapping (x). The output is included in the output of stacked layer’ residual function F(x). In the training of deep network with backpropagation, the gradient of error is propagated and calculated to the shallow layer. In deep layer, this error tends to be small till it eventually decreases. This is named the gradient reducing problem of deep networks. The issue could be resolved by residual learning.
The actual residual branch or unit l in the ResNet shows rectified linear unit (ReLU), weight, and batch normalization (BN). The output and input of a residual unit are estimated by
To finely adjust the parameters involved in the ResNet34 model, the COA is applied to it. COA was simulated as individual intelligence and sexual incentive of chimps from its group hunting that was varying from the other social predator. During the chimp colony, there are 4 kinds of chimps allowed to hunt models such as driver, barrier, chaser, and attackers. From every various capability, however, this diversity is essential to successful hunt. In the male chimps hunt superior to females. If they caught and killed, meal was distributed to every hunting party member and even bystander [
From the mathematical processes of group, driving, blocking, chasing, and attacking are resulting. In order to mathematically the drive as well as chase the prey was signified as the formulas:
To mathematically apply the performance of chimp, it can be considered as initial optimum solution accessible by attacker, driver, barrier, and chaser are optimum informed on the place of potential prey. Therefore, 4 of optimum solutions yet gained was saved and other chimps were required for updating its places based on optimum chimps places. This connection was written as subsequent formulas,
If the arbitrary values have been lying from the range of –1 and 1, the next place of chimp is from someplace among from present place and place of prey.
In the entire formulas:
The normal upgrading place process or the chaotic procedure for updating the place of chimps in optimization. The mathematical process was written as:
At the final stage, the WNN model is utilized to categorize the inputs into presence or absence of breast cancer. The wavelets are attained by scaling and translating a different function
WNN is non-linear regression framework which demonstrates input-output maps by relating wavelets with suitable scaling as well as translation. The output of WNN was defined as:
This section investigates the performance of the ADL-BCD technique interms of different dimensions [
The confusion matrices produced by the ADL-BCD technique take place under five distinct runs in
Methods | Precision | Recall | Specificity | Accuracy | F-Score |
---|---|---|---|---|---|
Training/Testing (90:10) | |||||
Normal | 0.9526 | 0.9617 | 0.9115 | 0.9441 | 0.9571 |
Benign | 0.8750 | 0.9180 | 0.9693 | 0.9596 | 0.8960 |
Malignant | 0.9787 | 0.8846 | 0.9963 | 0.9783 | 0.9293 |
Average | 0.9354 | 0.9215 | 0.9590 | 0.9607 | 0.9275 |
Training/Testing (80:20) | |||||
Normal | 0.9289 | 0.9378 | 0.8673 | 0.9130 | 0.9333 |
Benign | 0.7969 | 0.8361 | 0.9502 | 0.9286 | 0.8160 |
Malignant | 0.9362 | 0.8462 | 0.9889 | 0.9658 | 0.8889 |
Average | 0.8873 | 0.8733 | 0.9354 | 0.9358 | 0.8794 |
Training/Testing (70:30) | |||||
Normal | 0.9330 | 0.9330 | 0.8761 | 0.9130 | 0.9330 |
Benign | 0.7647 | 0.8525 | 0.9387 | 0.9224 | 0.8062 |
Malignant | 0.9333 | 0.8077 | 0.9889 | 0.9596 | 0.8660 |
Average | 0.8770 | 0.8644 | 0.9346 | 0.9317 | 0.8684 |
Training/Testing (60:40) | |||||
Normal | 0.9279 | 0.9234 | 0.8673 | 0.9037 | 0.9257 |
Benign | 0.7353 | 0.8197 | 0.9310 | 0.9099 | 0.7752 |
Malignant | 0.8478 | 0.75 | 0.9741 | 0.9379 | 0.7959 |
Average | 0.837 | 0.831 | 0.9241 | 0.9172 | 0.8323 |
Training/Testing (50:50) | |||||
Normal | 0.9261 | 0.8995 | 0.8673 | 0.8882 | 0.9126 |
Benign | 0.7042 | 0.8197 | 0.9195 | 0.9006 | 0.7576 |
Malignant | 0.8333 | 0.7692 | 0.9704 | 0.9379 | 0.8000 |
Average | 0.8212 | 0.8295 | 0.9191 | 0.9089 | 0.8234 |
A detailed comparison study of the ADL-BCD technique with distinct measures take place in
Methods | Precision | Recall | Specificity | Accuracy |
---|---|---|---|---|
MLP | 0.7800 | 0.7600 | 0.7400 | 0.7600 |
J48 + K-mean clustering | 0.9000 | 0.9000 | 0.8800 | 0.9000 |
Proposed DL technique | 0.9000 | 0.9300 | 0.9000 | 0.9200 |
AlexNet | 0.9189 | 0.9360 | 0.9178 | 0.9270 |
VGGNec | 0.9176 | 0.9358 | 0.9242 | 0.9278 |
GooleNet | 0.9224 | 0.9390 | 0.9317 | 0.9354 |
DenseNet | 0.9254 | 0.9459 | 0.9390 | 0.9387 |
DenseNet-II | 0.9285 | 0.9560 | 0.9536 | 0.9455 |
ADL-BCD | 0.9354 | 0.9215 | 0.9590 | 0.9607 |
Similarly, on examining the results with respect to recall, the comparative outcomes showcased that the MLP manner has attained ineffectual result with the recall of 0.76. Simultaneously, the J48 + K-mean Clustering and DL approaches have reached a recall of 0.9 and 0.93 correspondingly. Likewise, the VGGNec, AlexNet, GoogleNet, DenseNet, and DenseNet-II methods have attained a moderately closer recall of 0.9360, 0.9358, 0.9390, 0.9459, and 0.9560 correspondingly. But, the ADL-BCD methodology has resulted in a superior recall of 0.9215.
Furthermore, on investigating the performance with respect to specificity, the comparative outcomes outperformed that the MLP manner has reached ineffectual outcome with the specificity of 0.74. Also, the J48 + K-mean Clustering and DL algorithms have gained a specificity of 0.88 and 0.9 correspondingly. Similarly, the VGGNec, AlexNet, GoogleNet, DenseNet, and DenseNet-II methods have attained a moderately closer specificity of 0.9178, 0.9242, 0.9317, 0.9390, and 0.9536 correspondingly. But, the ADL-BCD methodology has resulted in a maximum specificity of 0.9590.
On exploratory the performance interms of accuracy, the comparative outcomes showcased that the MLP model has attained ineffectual outcome with the accuracy of 0.76. Simultaneously, the J48 + K-mean Clustering and DL manners have obtained an accuracy of 0.9 and 0.92 correspondingly. Besides, the VGGNec, AlexNet, GoogleNet, DenseNet, and DenseNet-II manners have reached a moderately closer accuracy of 0.9270, 0.9278, 0.9354, 0.9387, and 0.9455 correspondingly. Eventually, the ADL-BCD methodology has resulted in a superior accuracy of 0.9607. The simulation outcome indicated that the ADL-BCD model outperforms the state of art methods in terms of different measures.
This study has introduced a new ADL-BCD technique to detect the occurrence of breast cancer using digital mammograms. The goal of the ADL-BCD technique is to properly detect the existence of breast lesions using digital mammograms. The ADL-BCD technique contains GF based pre-processing, Tsallis entropy based segmentation, ResNet34 based feature extraction, COA based parameter tuning, and WNN based classification. The usage of COA based hyperparameter optimization helps to considerably boost the diagnostic efficiency. The ADL-BCD method is evaluated using a benchmark dataset and the results are analyzed under several performance measures. The simulation outcome indicated that the ADL-BCD model outperforms the state of art methods in terms of different measures. As a part of future scope, the advanced DL based instance segmentation technique can be designed to further increase the diagnostic outcome.