A Transfer Learning-Based Deep Learning Model for Automated Breast Cancer Identification in Mammograms

doi:10.21203/rs.3.rs-3749398/v1

Download PDF

Research Article

A Transfer Learning-Based Deep Learning Model for Automated Breast Cancer Identification in Mammograms

https://doi.org/10.21203/rs.3.rs-3749398/v1

This work is licensed under a CC BY 4.0 License

Version 1

posted

You are reading this older preprint version

Read the latest preprint version →

Breast cancer is a severe health issue that affects women all over the world, underscoring the need for reliable and effective screening techniques. The early detection, diagnosis, and treatment of breast cancer are made possible by computer-aided diagnostic (CAD) systems that rely on mammograms. This study introduces a unique deep-learning model that uses transfer learning to identify and categorize breast cancer automatically. Deep convolutional neural networks have been shown in several recent studies to diagnose breast cancer in mammograms with performance comparable to or even outperforming that of human experts. In order to extract features from the dataset from the Mammographic Image Analysis Society (MIAS), the proposed model uses pre-trained convolutional neural network (CNN) architectures like ResNet50 and Visual Geometry Group networks (VGG)-16. This novel deep-learning model holds significant potential for enhancing the efficiency and accuracy of breast cancer detection and classification.

Biomedical Engineering

Breast cancer

Screening techniques

Computer-aided diagnostic (CAD) systems

Mammograms

Transfer learning

Deep learning

Convolutional Neural Networks

Visual Geometry Group networks (VGG-16)

ResNet50

MIAS dataset

Cancer tumors are abnormal growths that invade surrounding tissues. There are two types of tumors: benign and malignant. Benign tumors are noncancerous and grow only locally, while malignant tumors consist of cancerous cells that can spread and invade other parts of the body. Breast cancer is a common malignant tumor affecting approximately 12% of Indian women during their lifespan. In India, a woman receives a breast cancer diagnosis every two minutes on average. Early detection through regular screening is essential for successful treatment, which may involve surgery, radiation therapy, chemotherapy, hormone therapy, and targeted therapy.

According to cancer statistics, Breast cancer is the most common type of cancer and has the highest frequency among women globally. The projected leading cancer cases of the breast would be 2,32,832 among females[1]. According to a study by the WHO[2], the prognosis and survival rate of breast cancer can be improved by early identification. Decades of research have focused on improving breast cancer diagnosis by reducing errors and speeding up the process. These efforts offer potential benefits for rapid diagnosis by radiologists and specialists. Using image processing and pattern recognition to automatically diagnose breast cancer from mammograms can mitigate human errors and expedite diagnosis. While the causes of breast cancer remain incompletely understood, prevention remains elusive. However, early detection significantly enhances the likelihood of a successful recovery, even though breast cancer often goes undetected until it reaches an advanced stage[3]-[4].

Recently, Machine learning has been used to create computer-aided detection (CAD) systems for identifying benign and malignant lesions in image processing, artificial intelligence, and neural network approaches. These technologies can improve image quality, automate image readability, and reduce experimental errors for more accurate diagnosis[5]-[7]. Machine learning (ML) and artificial intelligence (AI) approaches are being used in many research publications that examine the detection, segmentation, and classification of breast cancer. This research frequently emphasizes binary classification tasks in particular, broadening their applicability beyond breast cancer to cover a variety of different cancer types, such as lung, brain, skin, stomach, and kidney cancer. These innovations have the potential to significantly improve the precision and efficacy of cancer detection and therapy[8]-[12].

Deep learning, a more modern technique, has made it possible to effectively address traditional machine learning problems. In particular, for tasks like image classification and object localization, deep learning excels at producing excellent feature representations. Convolutional Neural Networks (CNNs) stand out among the extensively used deep learning techniques. CNNs are specifically designed for 2D image data, and their architecture has been improved to support this structure. However, it's important to keep in mind that effective CNN training necessitates access to a sizable number of data, which is sometimes missing in the medical community, particularly in the context of breast cancer research[13]-[15].

The primary contribution of this study is to classify breast cancer from mammographic images using a deep convolutional neural network through the following steps:

Accelerating training in breast image analysis involves focusing on critical regions indicating abnormalities.
Enhancing classification by modifying the pre-trained network's classifier to improve performance and accuracy.
Addressing overfitting concerns.
Employing a comprehensive set of six evaluation metrics: AUC, F-score, sensitivity, specificity, precision, and accuracy.

The manuscript is organized as follows for the remaining portions: Section 2 gives a background on computer-aided diagnostics and breast cancer. Section 3 contains the literature review. Section 4 provides details about the suggested methodology. In Section 5, the experiment's findings and analysis are described. Our task is finished in Section 6.

Yan et al. [16] developed a computer-aided detection model for breast cancer. Their strategy entailed creating a powerful algorithm and collecting a carefully labeled dataset for model training. The authors present a feature-sensitive deep convolutional neural network for breast cancer diagnosis that makes use of an end-to-end training process. The first step of the procedure is to extract important image characteristics using a pre-trained model enhanced with custom layers. It then includes a feature fusion module that determines the weighting of each feature vector. By allowing various occurrences of each case to have varied degrees of effect on the classifier, this adaptive feature weighting technique ultimately improves the accuracy of breast cancer identification.

Xiang Yu et al. [17] implemented a new data augmentation methodology called SCDA (Scaling and Contrast Limited Adaptive Histogram Equalization Data Augmentation). This can modify the accuracy and robustness of machine learning models by introducing more variation in the data. Overall, SCDA is a new technique for data augmentation that can improve the performance of machine learning models, particularly in computer vision tasks that involve image processing.

Xin Shu et al. [18] developed an innovative approach for classifying mammograms without the need for mask ground truth labels. Convolutional neural networks (CNNs) are used in their end-to-end full-image mammography classification technique, It perform better than previous cutting-edge classifiers and detection techniques reliant on segmentation annotations. This breakthrough has the potential to improve mammographic image analysis and enhance breast cancer detection.

Ritabrata Sanyal et al. [19] presented a Hybrid Ensemble Framework for Patch-wise Classification of High-Resolution Breast Histopathology Images using Multiple Fine-tuned CNNs and XGBoot. In this system, optimised convolutional neural network (CNN) designs are integrated with extreme gradient boosting trees (XGBoost), the top-level classifier, to be used as supervised feature extractors. Experimental results show that the proposed strategy outperforms the state-of-the-art methods in patch-wise classification of high-resolution breast histopathology images.

Q. Abbas [20] introduces DeepCAD, a cutting-edge CAD solution that uses a four-phase technique to solve a variety of problems. This technique retrieves descriptors using measures of Local Binary Pattern Variance (LBPV) and Speed-Up Robust Features (SURF). These descriptors are subsequently converted into invariant features, assuring their stability and dependability for additional research. Deep Invariant Features (DIFs) utilize a multilayer deep-learning architecture by DeepCAD in both supervised and unsupervised ways to further improve performance and attain sensitivity, specificity, accuracy, and AUC of 92%, 84.2%, 91.5%, and 0.91, respectively.

Sha et al. [21] represented a Grasshopper Optimisation Algorithm for image noise reduction. Convolutional neural network image segmentation that is optimized feature extraction and selection using the same algorithm. The suggested method offers a practical solution for cancer area localization in mammography images by enhancing precision and lowering computing costs using these strategies. The suggested technique tries to automatically identify and categorize malignant areas in breast pictures. With sensitivity, specificity, and accuracy rates of 96%, 93%, and 92%, respectively, this strategy performed admirably.

Sheetal Rajpal et al. [22], proposed a deep-learning-based ground-breaking framework for the detection of biomarkers and the diagnosis of breast cancer subtypes. This framework consists of three stages: To obtain a compact representation of the gene expression data in the first stage, an autoencoder is utilized. The second stage involves categorizing breast cancer subtypes using a supervised feed-forward neural network. An algorithm called the Biomarker Gene Discovery Algorithm (BGDA) is suggested in the third phase to determine the significance of various genes. the framework obtained a mean accuracy of 0.899 with a confidence interval of 0.04 using 10-fold cross-validation.

Mondol et al. [23] used an adversarial auto-encoder (AAE) and a dual-stage neural network architecture called AFExNet, and they were able to extract features from highly dimensional genetic data. Using a publicly available RNA-Seq dataset of breast cancer, They tested the effectiveness of their model using twelve alternative supervised classifiers. While AFExNet consistently outperforms 12 other classifiers across all performance parameters, their model classifier is independent of AFExNet. They also created a technique called "TopGene" to locate highly weighted genes in the latent space, which may help locate cancer biomarkers. AFExNet offers a lot of potential for properly and efficiently extracting features from biological data.

Guangli Li et al [24]proposed a new model called Multi-View Attention-Guided Multiple Instance Detection Networks, dividing each histopathology image into instances to use high-resolution information fully. The Multiple-View Attention (MVA) approach was established by the algorithm to locate and localise lesion patches within the image by focusing on certain instances. Combining this with a Multiple Instance Learning (MIL) pooling technique driven by MVA allowed instance-level characteristics to be aggregated for subsequent bag-level classification. The model achieved better localization results without sacrificing classification performance, making it highly practical.

Amin Ul Haq et al. [25] proposed supervised and unsupervised techniques for related feature selection from a data set and determined by PCA algorithms and the Relief algorithms. These features are closely connected with accurate breast cancer diagnosis. The suggested approach obtained 99.91% accuracy on features chosen by the Relief algorithm, achieving excellent results in terms of accuracy.

Yasin Yari et al. [26] proposed deep transfer learning models that leverage pre-trained DCNNs to enhance binary and multiclass classification in breast cancer detection. By utilizing pre-trained weights from ResNet50 and DenseNet on the ImageNet dataset, the models were fine-tuned with a deep classifier and data augmentation techniques to identify malignant and benign tissues. The system achieved impressive accuracies of up to 98% in multiclass classification and up to 100% in binary classification.

Zhiqiong Wang et al. [27] developed a technique for identifying breast masses using CNN deep features in combination with unsupervised Extreme Learning Machine (ELM) clustering. By combining deep features with morphological, texture, and density characteristics, they subsequently created a complete feature set. Finally, the suggested technique for mass detection and breast cancer classification has undergone extensive testing to demonstrate its accuracy and effectiveness.

Irum Hirra et al. [28] introduced Pa-DBN-BC, a novel approach for breast cancer detection and classification in histopathology images. Their method, based on Deep Belief Networks (DBNs), automatically extracts features from image patches using unsupervised pre-training and supervised fine-tuning. The proposed approach has the potential to enhance breast cancer analysis, as evidenced by its accuracy of 86% on a variety of datasets, which is better than both earlier deep learning techniques and regular methods.

Farnoos Azoor et al. [29] combined efficientNet with other pre-trained models, they enhanced accuracy by using pre-trained cnn-based models notably with greater performance and less number of parameters, and the network resilience was increased much further using ensemble learning their tests showed encouraging test accuracy results of 9605 for abnormality type classification and 8571 for pathology diagnostic classification using 10-fold cross-validation this study demonstrates how pre-trained models and ensemble learning are beneficial for correctly classifying medical images.

Daniel et al. [30] presented an efficientNet-based convolutional network on the CBIS-DDSM dataset. The suggested model attained an AUC of 0.93 and an accuracy of 85.13% with the use of the 5-fold cross-validation approach. In comparison to VGG and Resnet, they found that modern efficientNet based model performed well.

H. U. Khan et al. [31]created a breast cancer classification algorithm for histopathology images that makes use of a multi-scale feature fusion strategy. Their model is made up of two blocks, each of which has three lightweight sub-models, for a total of six sub-models. This architecture is made to gather data from input images at different dimensions, improving its capacity to correctly categorize cases of breast cancer. The predictive performance evaluated using different performance measures shows that this model produces a better result than already existing approaches.

In this section, we discuss the data and architecture of the network that we have designed for breast cancer classification. Figure 1 shows the schematic of the system. Our breast cancer detection scheme involves the following stages: 1) data acquisition, 2) preprocessing, and 3) deep CNN training based on transfer learning.

3.1 Dataset Description

In this study, two publicly available datasets of raw and scanned mammography data that are recorded in DICOM format are used to validate the two approaches that are suggested. The two datasets are described in detail below. CBIS-DDSM [32]The CBIS-DDMS (Curated Breast Imaging Subset of DDSM) comprises decompressed images and exact annotations, packaged similarly to contemporary computer vision data sets. There are 753 calcification instances, 891 mass cases, and 3071 images in the CBIS-DDMS. The work mentioned used a mammogram database provided by MIAS[33]. The database consists of 322 images in the portable gray map (PGM) format, each with a size of 1024 x 1024 pixels. The images are categorized into three classes: benign (62 images), malignant (51 images), and normal (209 images). Additional details about the data are presented in Table 1.

Table 1

Details of MINI-MIAS
Class	Sub-class of abnormality present	Numbers	Total
Benign	Well-defined Circumscribed Masses(CIRAC) Calcination(CALC) Spiculated Masses(SPIC) Ill-defined Masses(MISC) Architectural Distortion(ARCH) Asymmetry(ASYM)	19 10 11 7 9 6	62
Malignant	Well-defined Circumscribed Masses(CIRAC) Calcination(CALC) Spiculated Masses(SPIC) Ill-defined Masses(MISC) Architectural Distortion(ARCH) Asymmetry(ASYM)	4 13 8 7 10 9	51
Normal	-	-	209

3.2 Data Preprocessing

Data preprocessing plays a vital role in training Convolutional Neural Networks (CNNs), which are widely used for image classification, object detection, and other computer vision tasks. Here are some key considerations for data preprocessing in CNN architecture:

3.2.1 Image Resizing

CNN models normally assume that input images have a fixed size. Therefore, resizing the images in the dataset to a consistent shape is essential before feeding them into the network. Commonly used sizes are square images like 224x224 or 256x256 pixels.

3.2.2 Data Normalization

Normalizing the pixel values of images is crucial to ensure that the data falls within a similar range. This helps the CNN model converge faster during training. The values of the pixels are typically scaled to a range between 0 and 1, or mean subtraction and standard deviation scaling are used.

3.2.3 Data Augmentation

Data augmentation helps the CNN model generalize better and reduces overfitting by providing more diverse examples for training[34]. Large datasets are preferable for use with DL models. This technique artificially increases the training dataset's size by applying random transformations to the images, such as rotation, translation, scaling, flipping, or introducing random noise.

3.2.4 Data Splitting

The dataset must be divided into training, validation, and testing sets. The validation set is used for hyperparameter tuning and model selection, the testing set is used to assess the final model's performance on untrained data, and the training set is used to train the CNN model. The MIAS dataset is split into a training set (80%) and a testing set (20%), facilitating robust evaluation and validation of models for breast cancer classification.

3.3 Deep-CNN training based on transfer learning

In this study, we constructed a variety of CNN architectures for training and feature extraction. We used VGG16 and ResNet networks for feature extraction.

3.3.1 VGG16

VGG16 is a popular Convolutional Neural Network (CNN) architecture that was introduced by the Visual Geometry Group (VGG) at the University of Oxford. It has achieved remarkable performance on image classification tasks and serves as a strong baseline for various computer vision tasks. The basic architecture of VGG16 is shown in Fig. 6. It has 16 convolutional layers, including 13 blocks and 3 fully connected layers. There are three convolutional blocks that are stacked on top of one another, maintaining a small filter size of 3x3. In max pooling layers[35], the number of filters gradually increases while spatial dimensions decrease. In VGG16, each convolutional block consists of multiple convolutional layers followed by a max pooling layer. Each convolutional layer has 64, 128, 256, and 512 filters, progressively increasing the depth of the network. After each convolutional layer, ReLU activation functions are used to introduce non-linearity to the network, allowing it to learn complex representations. Max pooling with a 2x2 filter and a stride of 2 is applied after each set of convolutional layers to reduce the spatial dimensions. VGG16 includes three fully connected layers with 4096 units each, followed by a softmax layer for classification. These layers integrate the spatial information learned by the convolutional layers and produce class probabilities. VGG16 is typically trained on large-scale datasets like ImageNet, where it learns generic features from a vast number of images. After pretraining, the fully connected layers are often fine-tuned on a specific task or dataset[35].

3.3.2 ResNet50

The ResNet (Residual Network) family includes the widely used ResNet-50 Convolutional Neural Network (CNN) architecture. It was developed by Microsoft Research and is very effective at many computer vision tasks[36]. The basic ResNet50 architecture is shown in Fig. 6. ResNet-50 is an architecture built upon the Residual Network (ResNet) concept, which addresses the degradation issue when training deep neural networks. There are 50 layers total, including fully connected, pooling, convolutional, and shortcut connections (residual connections).

The input convolutional layer applies a 7x7 kernel size convolution operation to the input image, followed by batch normalization, ReLU activation, and another ReLU activation. Downsampling of the input is performed in this layer. To further reduce the spatial dimensions, a max pooling layer with a 3x3 kernel and a stride of 2 is used after the input convolutional layer. The final layer of ResNet-50 is a fully connected layer with a softmax activation function. It takes the vector representation from the global average pooling layer and performs a matrix multiplication with the weight matrix of the fully connected layer, providing the class probabilities for image classification. During training, the model is fed with input images in small batches. The images are processed through the model, and the loss is computed based on the expected results. Gradients are then backpropagated to update the model's parameters. This process is repeated for multiple epochs, representing complete passes through the training dataset.

4.1 Environment

The experiment involved training and testing of the proposed CNN are done on a computer with intel core i5, 11G2080Ti Graphics Processing Unit(GPU), 16GB memory, windows 10 operating system, pytorch framework, and python 3.7. The training parameters are set as 16 batch size, 0.001 learning rate, Adam optimization, and 50 epochs.

4.2 Experimental Results

This section presents a series of trials carried out to examine the performance of the suggested model on the MIAS dataset. In this study, two different deep learning models, VGG16 and ResNet50, were trained using transfer learning (TL). The study analyzed the output from several models in terms of AUC, sensitivity, specificity, accuracy, and precision. Three categories were used to aggregate the data: "Benign," "Malignant," and "Normal." It was then split into two portions: 20% for testing and 80% for training. The performance of the three classes was compared using a set of metrics to determine how effective the suggested model was, as shown in Table 2. The assessment included a number of factors that allowed for the evaluation of various approaches based on True Positive (TP), False Positive (FP), False Negative (FN), and True Negative (TN), which are defined as follows:

True Positive (TP): Instances correctly classified as positive in the "Benign," "Malignant," or "Normal" categories.
False Positive (FP): Instances incorrectly classified as positive when they are not.
False Negative (FN): Instances incorrectly classified as negative when they are positive.
True Negative (TN): Instances correctly classified as negative in the respective categories.

These parameters collectively allowed for a comprehensive evaluation of the model's performance across the different classes and categories.

Table 2

Evaluation metrics of the proposed model
Classes	Benign	Malignant	Normal
Benign	P_BB	P_MB	P_NB
Malignant	P_BM	P_MM	P_NM
Normal	P_BN	P_MN	P_NN

4.2.1 Accuracy

Accuracy is a widely used evaluation metric that measures the proportion of cases that are properly identified out of all occurrences, which is defined as

Accuracy = \(\frac{P{BB}_{}+PBB+PBB}{PMB+PNB+PBM+PNM+PBN+PMN}\) (1)

VGG16 has achieved an astonishing 97.02% accuracy rate, demonstrating excellent performance.

4.2.2 Precision

The percentage of true positives (positive samples that were properly predicted to be positive) in comparison to the total number of positive predictions (true positives plus false positives) is measured by the precision metric. It acts as a metric for how well the model can recognize positive samples.

With an accuracy rate of 34.84%, the ResNet50 model surpasses other models. This suggests that, as compared to the other models being tested, ResNet50 is especially skilled at delivering positive predictions with a greater level of accuracy.

4.2.3 Sensitivity

Sensitivity, also called recall or true positive rate, measures the model's accuracy in identifying all real positive cases as positive occurrences. A higher sensitivity value indicates a lower rate of false negatives, indicating a more effective model in capturing positive instances. With a sensitivity score of 55.76%, the VGG16 obtains the best performance.

4.2.4 Specificity

A measure of a model's precision in accurately recognising negative occurrences among all real negative situations is called specificity. A higher specificity value indicates a lower rate of false positives, demonstrating a more accurate model in identifying negative instances. VGG16 attains an impressive specificity rate of 69.68%, showcasing its ability to accurately identify true negatives.

4.2.5 AUC

AUC (Area Under the Curve) is a widely used metric in binary classification tasks. The area under the ROC curve is used to assess a model's general performance. A higher AUC value denotes a model's ability to discriminate between positive and negative.

4.3 Analysis of Results

In Table 3, the outcomes of a 10-fold cross-validation are presented. It can be noticed that the cross-validation method outperformed the 80 − 20 technique in both CNNs. Table 4 displays the Breast Cancer (BC) classification performance of several Convolutional Neural Networks (CNNs) employing the 80:20 method indicating the evaluation of these models using an 80% training set and a 20% testing set. Table 5, which compares the performance with four other existing models, displays the outcomes of the studies. In terms of accuracy, sensitivity, specificity, and AUC, the suggested model performs better than other current models, according to the results of the examination.

Table 3

Breast Cancer classification performance of various CNNs after preprocessing using 10-fold cross-validation.
CNN	Classification Performance
CNN	Accuracy (%)	Sensitivity(%)	Specificity (%)	Precision (%)	AUC	F-score (%)
VGG16	97.02	95.44	96.65	97.00	0.96	96.99
ResNet50	95.71	94.73	95.60	95.72	0.95	95.70

Table 4

CNN Performance Analysis with 80 − 20 Technique Preprocessing
CNN	Classification Performance
CNN	Accuracy (%)	Sensitivity(%)	Specificity (%)	Precision (%)	AUC	F-score (%)
VGG16	96.00	94.40	95.65	96.01	0.95	95.07
ResNet50	94.70	93.70	94.60	94.71	0.94	94.60

Table 5

The proposed model is compared to the existing models.
Method	Classification Performance
Method	Accuracy (%)	Sensitivity (%)	Specificity (%)	Precision (%)	AUC	F-score (%)
Daniel et al.	85.13	-	-	-	0.83	-
Irum Hirra et al.	86.00	87.09	84.00	-	-	-
Proposed	96.00	94.40	95.65	96.01	0.95	95.07

Resnet50

In order to increase the classification accuracy of the images in the MIAS dataset, a unique deep-learning model is presented in this study. This model's primary objective is to aid doctors in the early detection and diagnosis of breast cancer. Images from the MIAS dataset are divided into three groups: benign, malignant, and normal. The original MIAS dataset underwent a number of preprocessing processes before analysis, including noise reduction, enhancing breast image contrast, removing non-breast regions, and identifying malignant cancer. On the basis of an in-depth review of several deep convolutional neural network techniques, we developed a very efficient architecture employing pre-trained VGG16 and ResNet50 models combined with a novel fully-connected classifier. Our VGG16 model achieved 96.00% accuracy, 0.95 AUC, 94.40% sensitivity, 95.65% specificity, 96.01% precision, and 96.99% F-score. These results are better than the other mentioned methods.

Sathishkumar, K., Chaturvedi, M., Das, P., Stephen, S., & Mathur, P. (2023). Cancer incidence estimates for 2022 & projection for 2025: Result from National Cancer Registry Programme, India. Indian Journal of Medical Research, 0(0), 0. https://doi.org/10.4103/ijmr.ijmr_1821_22
Globocan 2020: India Factsheet. (2021). The Global Cancer Observatory. Updated March 2021. https:// gco.iarc.fr/today/data/factsheets/populations/356-india-fact-sheets.pdf. Retrieved December 3, 2021.
Mondol, R. K., Truong, N. D., Reza, M., Ippolito, S., Ebrahimie, E., & Kavehei, O.. (2021). AFExNet: An Adversarial Autoencoder for Differentiating Breast Cancer Sub-types and Extracting Biologically Relevant Genes. 19(4), pp. 2060–2070. https://doi.org/DOI 10.1109/TCBB.2021.3066086.
H. Li, D. Chen, W. H. Nailon, M. E. Davies and D. I. Laurenson, "Dual Convolutional Neural Networks for Breast Mass Segmentation and Diagnosis in Mammography," in IEEE Transactions on Medical Imaging, vol. 41, no. 1, pp. 3-13, Jan. 2022, doi: 10.1109/TMI.2021.3102622.
S.C. Tai, Z.-S. Chen, and W.-T. Tsai, ``An automatic mass detection system in mammograms based on complex texture features,'' IEEE J. Biomed. Health Inform., vol. 18, no. 2, pp. 618627, Mar. 2014.
Azour, F., & Boukerche, A. (2022). An Efficient Transfer and Ensemble Learning based Computer Aided Breast Abnormality Diagnosis System. IEEE Access, 1–1. https://doi.org/10.1109/ACCESS.2022.3192857.
M. Kowal et al., “Computer-aided diagnosis of breast cancer based on fine needle biopsy microscopic images”, Computers in Biology and Medicine, vol. 43 (10), pp. 1563–1572, 2013.
W. Zhu, C. Ma, X. Zhao, M. Wang, A. A. Heidari, H. Chen, and C. Li,``Evaluation of sino foreign cooperative education project using orthogonal sine cosine optimized kernel extreme learning machine,'' IEEE Access, vol. 8, pp. 61107-61123, 2020.
J. Tu, A. Lin, H. Chen, Y. Li, and C. Li, ``Predict the entrepreneurial intention of fresh graduate students based on an adaptive support vector machine framework,'' Math. Problems Eng., vol. 2019, pp. 1-16, Jan. 2019.
A. Lin, Q. Wu, A. A. Heidari, Y. Xu, H. Chen, W. Geng, and C. Li, ``Predicting intentions of students for master programs using a chaos-induced sine cosine-based fuzzy K-nearest neighbor classier,'' IEEE Access, vol. 7, pp. 67235-67248, 2019.
Y. Wei, N. Ni, D. Liu, H. Chen, M. Wang, Q. Li, X. Cui, and H. Ye, ``An improved grey wolf optimization strategy enhanced SVM and its application in predicting the second major,'' Math. Problems Eng., vol. 2017, pp. 1-12, Feb. 2017.
Y. Xu, K. Yu, P. Wang, H. Chen, X. Zhao, and J. Zhu, ``A new hybrid machine learning approach for prediction of phenanthrene toxicity on mice,'' IEEE Access, vol. 7, pp. 138461-138472, 2019.
Q. Zhang, L. T. Yang, Z. Chen, and P. Li, ``A survey on deep learning for big data,'' Inf. Fusion, vol. 42, pp. 146-157, Jul. 2018.
N. Dhungel, G. Carneiro, and A. P. Bradley, ``A deep learning approach for the analysis of masses in mammograms with minimal user intervention,'' Med. Image Anal., vol. 37, pp. 114-128, Apr. 2017.
P. Xi, C. Shu, and R. Goubran, ``Abnormality detection in mammography using deep convolutional neural networks,'' in Proc. IEEE Int. Symp. Med. Meas. Appl. (MeMeA), Jun. 2018, pp. 1-6.
Wang, Y., Zhang, L., Shu, X., Feng, Y., Yi, Z., & Lv, Q. (2022). Feature-Sensitive Deep Convolutional Neural Network for Multi-Instance Breast Cancer Detection. IEEE/ACM Transactions on Computational Biology and Bioinformatics, 19(4), 2241–2251. https://doi.org/10.1109/TCBB.2021.3060183.
Yu, X., Kang, C., Guttery, D. S., Kadry, S., Chen, Y., & Zhang, Y.-D. (2021). ResNet-SCDA-50 for Breast Abnormality Classification. IEEE/ACM Transactions on Computational Biology and Bioinformatics, 18(1), 94–102. https://doi.org/10.1109/TCBB.2020.2986544.
Shu, X., Zhang, L., Wang, Z., Lv, Q., & Yi, Z. (2020). Deep Neural Networks With Region-Based Pooling Structures for Mammographic Image Classification. IEEE Transactions on Medical Imaging, 39(6), 2246–2255. https://doi.org/10.1109/TMI.2020.2968397.
Sanyal, R., Kar, D., & Sarkar, R. (2022). Carcinoma Type Classification From High-Resolution Breast Microscopy Images Using a Hybrid Ensemble of Deep Convolutional Features and Gradient Boosting Trees Classifiers. IEEE/ACM Transactions on Computational Biology and Bioinformatics, 19(4), 2124–2136. https://doi.org/10.1109/TCBB.2021.3071022.
Q. Abbas, ``DeepCAD: A computer-aided diagnosis system for mammographic masses using deep invariant features,'' Computers, vol. 5, no. 4, p. 28, Oct. 2016, doi: 10.3390/computers5040028.
Sha, Zijun, Lin Hu, and Babak Daneshvar Rouyendegh. 2020. “Deep Learning and Optimization Algorithms for Automatic Breast Cancer Detection.” International Journal of Imaging Systems and Technology 30 (2): 495–506. https://doi.org/10.1002/ima.22400.
Rajpal, S., Agarwal, M., Kumar, V., Gupta, A., & Kumar, N. (2021). Triphasic DeepBRCA-A Deep Learning-Based Framework for Identification of Biomarkers for Breast Cancer Stratification. IEEE Access, 9, 103347–103364. https://doi.org/10.1109/ACCESS.2021.3093616
Mondol, R. K., Truong, N. D., Reza, M., Ippolito, S., Ebrahimie, E., & Kavehei, O. (2022). AFExNet: An Adversarial Autoencoder for Differentiating Breast Cancer Sub-Types and Extracting Biologically Relevant Genes. IEEE/ACM Transactions on Computational Biology and Bioinformatics, 19(4), 2060–2070. https://doi.org/10.1109/TCBB.2021.3066086
Li, G., Li, C., Wu, G., Ji, D., & Zhang, H. (2021). Multi-View Attention-Guided Multiple Instance Detection Network for Interpretable Breast Cancer Histopathological Image Diagnosis. IEEE Access, 9, 79671–79684. https://doi.org/10.1109/ACCESS.2021.3084360
Haq, A. U., Li, J. P., Saboor, A., Khan, J., Wali, S., Ahmad, S., Ali, A., Khan, G. A., & Zhou, W. (2021). Detection of Breast Cancer Through Clinical Data Using Supervised and Unsupervised Feature Selection Techniques. IEEE Access, 9, 22090–22105. https://doi.org/10.1109/ACCESS.2021.3055806
Y. Yari, T. V. Nguyen, and H. T. Nguyen, "Deep Learning Applied for Histological Diagnosis of Breast Cancer," in IEEE Access, vol. 8, pp. 162432-162448, 2020, doi: 10.1109/ACCESS.2020.3021557.
Wang, Zhiqiong, Mo Li, Huaxia Wang, Han-Yu Jiang, Yu-Dong Yao, Hao Zhang, and Junchang Xin. 2019. “Breast Cancer Detection Using Extreme Learning Machine Based on Feature Fusion with CNN Deep Features.” IEEE Access 7 (January): 105146–58. https://doi.org/10.1109/access.2019.2892795.
Hirra, I., Ahmad, M., Hussain, A., Ashraf, M. U., Saeed, I. A., Qadri, S. F., Alghamdi, A. M., & Alfakeeh, A. S. (2021). Breast Cancer Classification From Histopathological Images Using Patch-Based Deep Learning Modeling. IEEE Access, 9, 24273–24287. https://doi.org/10.1109/ACCESS.2021.3056516.
F. Azour and A. Boukerche, "An Efficient Transfer and Ensemble Learning Based Computer Aided Breast Abnormality Diagnosis System," in IEEE Access, vol. 11, pp. 21199-21209, 2023, doi: 10.1109/ACCESS.2022.3192857.
D. G. P. Petrini, C. Shimizu, R. A. Roela, G. V. Valente, M. A. A. K. Folgueira and H. Y. Kim, "Breast Cancer Diagnosis in Two-View Mammography Using End-to-End Trained EfficientNet-Based Convolutional Network," in IEEE Access, vol. 10, pp. 77723-77731, 2022, doi: 10.1109/ACCESS.2022.3193250.
H. U. Khan, B. Raza, A. Waheed and H. Shah, "MSF-Model: Multi-Scale Feature Fusion-Based Domain Adaptive Model for Breast Cancer Classification of Histopathology Images," in IEEE Access, vol. 10, pp. 122530-122547, 2022, doi: 10.1109/ACCESS.2022.3223870.
N. Dhungel, G. Carneiro, and A. P. Bradley, “A deep learning approach for the analysis of masses in mammograms with minimal user intervention,” Medical Image Analysis, vol. 37, pp. 114–128, 2017. [Online]. Available: https://www.sciencedirect.com/science/article/pii/S136184151730018X
S. B. Y. Tasdemir, K. Tasdemir, and Z. Aydin, ``A review of mammographic region of interest classification,'' WIREs Data Mining Knowl. Discovery, vol. 10, no. 5, p. e1357, Sep. 2020, doi: 10.1002/widm.1357.
A. Krizhevsky, I. Sutskever, G.E. Hinton, Imagenet classification with deep convolutional neural networks, in: Advances in neural information processing systems, 2012, pp. 1097–1105.
K. Simonyan and A. Zisserman, ``Very deep convolutional networks for large-scale image recognition,'' 2014, arXiv:1409.1556. [Online]. Available: http://arxiv.org/abs/1409.1556.
K. He, X. Zhang, S. Ren, and J. Sun, ``Deep residual learning for image recognition,'' in Proc. IEEE Conf. Comput. Vis. Pattern Recognit. (CVPR), Jun. 2016, pp. 770778, doi: 10.1109/cvpr.2016.90.

The authors declare no competing interests.

Download PDF

Version 1

posted

You are reading this older preprint version

Read the latest preprint version →

A Transfer Learning-Based Deep Learning Model for Automated Breast Cancer Identification in Mammograms

Status:

Version 1

Abstract

Figures

1. INTRODUCTION

2. RELATED WORK

3. MATERIALS AND METHODS

3.1 Dataset Description

3.2 Data Preprocessing

3.2.1 Image Resizing

3.2.2 Data Normalization

3.2.3 Data Augmentation

3.2.4 Data Splitting

3.3 Deep-CNN training based on transfer learning

3.3.1 VGG16

3.3.2 ResNet50

4. EXPERIMENTS AND RESULTS

4.1 Environment

4.2 Experimental Results

4.2.1 Accuracy

4.2.2 Precision

4.2.3 Sensitivity

4.2.4 Specificity

4.2.5 AUC

4.3 Analysis of Results

5. Conclusion

References

Additional Declarations

Status:

Version 1