Breast Cancer Segmentation with Preprocessing and DNN

Breast Cancer Segmentation with Preprocessing and DNN

T Anil Raju

Associate Professor, ECE Dept.

Lakireddy Bali Reddy College of Engineering (Autonomous)

Mylavaram, India

1. Gayathri

   B.Tech Student, ECE Dept.

   Lakireddy Bali Reddy College of Engineering (Autonomous)

   Mylavaram, India

   V. Jayanth

   B.Tech Student, ECE Dept.

   Lakireddy Bali Reddy College of Engineering (Autonomous)

   Mylavaram, India

2. Pawan Kalyan

B.Tech Student, ECE Dept.

Lakireddy Bali Reddy College of Engineering (Autonomous)

Mylavaram, India

AbstractOne of the most crucial concerns that has to beaddressed globally is early breast cancer identification becauseit can assist patients have a higher survival percentage. Breast cancer can be discovered early with the help of mammograms, which can significantly lower the cost of treatment. Segmentationtechniques are necessary for the detection of breast tumors. In image analysis, segmentation which comprises detection, feature extraction, classification, and treatment plays a crucial role. Physicians use segmentation to calculate the amount of breast tissue for planning treatments Unsupervised machine learning techniques are more commonly employed in machine learning segmentation(U-Net) is typically made for mammography image- segmentation since it requires fewer annotated images than otherdeep-learning models. With no prior or subsequent processing,a deep learning model can be trained. The U-Net model willenhance computing when there are powerful GPUs present, facilitating the training of networks with additional layers. How- ever, recent studies have shown that early use of pre processing techniques into DNN would undoubtedly increase accuracy.

Index Terms:- Mammograms; Pectoral Muscles; Region of Interest; Segmentation.

1. INTRODUCTION

   One of the top causes of death for women globally is breast cancer. Early detection and diagnosis improve prognosisand decrease mortality. The World Health Organization claimsthat (WHO). The mortality rate, the expense of care, and the quality of life for patients can all be considerably reducedif it is discovered early on since a biopsyis not required. Breast cancer mortality is decreased by 25.0 percent and treatment choices are increased with early diagnosis. The identification, extraction, and categorization of breast tumors using mammography image segmentation were the subjects of a number of past studies that examined in this work, which discovered that mammography segmentation typically used kmeans segmentation, serves as an example. However, no quantitative statistics have been offered. Additionally, unsu- pervised machine learning is the foundation of k-means. K- means base is unsupervised machine learning as well. We also read over the papers.

   First, we provide a descriptionof the process used to segment mammogram images. We then discuss the most well-liked noise-reduction filters for mammography images. Third, we address the segmentation metrics and categorization of publicly and privately available mammography.

   Then examine the most well-liked deep learning, machine learning, and traditional segmentation techniques for data segmentation. Effective preprocessing of mammographic pictures is crucial before creating an intelligent system. This entails eliminating the background, adding noise, and deleting the pectoral muscle in addition to applying image enhancements. Researchers have previously proposed background and pec- toral muscle removal as two methods for picture segmentation,but little study has been made on methods for image en-hancement during the preprocessing stage. The main objectiveof this work is to present efficient image enhancement and segmentation techniques.Why segments are used in mammography images By recognising the masses in mammograms, segmentation in image processing involves cutting an image into numerous segments in order to retrieve the ROI from the image. Anomalies are very easy to locate. Pectoral muscles, onthe other hand, have the potential to obstruct identification; as a result, they should be removed before segmentation. Why segments are used in mammography images in image processing,segmentation entails dividing an image into numerous segments in order to identify the masses in mammograms and extract the ROI from the picture.

   Finding anomalies is extremely simple.Contrarily, pectoral muscles may prevent identification; as a result, they mustto be cut off before to segmentation. In order to enhancethe quality of noisy photos, noise and other local imperfections must be removed using filtering techniques. To obtain ROIs with possible masses, segmentation involves cutting the mammography image into many, nonoverlapping portions. However, a variety of circumstances, as discussed in, canmake mammography segmentation methods less effective andmake it harder to spot abnormalities in mammogram images. The 14 texture qualities used to categorise mammographymasses

   are insufficient for classifying ribbons or edges of masses with pixel resolutions finer than 200m or coarser than 800m per pixel. As opposed to this, the author found that ribbons of masses with pixel sizes of 400 and 800 m are the most useful when used with a Bayesian classifier based on mammography mass categorization. Mammogram images with higher pixel resolution require more computing to analyse, whereas methods for texture analysis may perform worse withtoo low of a pixel resolution. For this reason, the highest mammography pixel resolution should be employed.

   By changing the grey levels of the mammogram pictures, preprocessing techniques and feature normalisation may lessen the usefulness of texture analysis techniques, and feature normalisation may lessen classification accuracy.The texture feature has to be normalised in order to avoid smaller numeric ranges from dominating larger numeric ranges. Radiologists can swiftly identify breast cancer because to segmentation because benign and malignant tumour types differ from one another. The former favour regular shapes, whilst the latter frequently adopt irregular ones. Segmentation must come before feature extraction as a vital step. Preprocessings primary objective is to change an image so that the results are more appropriate for a certain application than the original. After segmentation, which can be completed by creating the images grey-level co-occurrence matrix and utilising GLCM features, automated mammography segmentation, the ROI is then utilised to extract features from the picture. There has been very little to no (semiautomatic) human interaction,as shown by the automated photo analysis and segmentation. Global thresholding, a technique used to segment images for a long time, has purportedly led to the misclassification of breast tumours. Recent research has demonstrated that mam- mographyinterpretations byradiologists commonly resulted in false-negative cases. Mammograms should be double-checked to reduce this inaccuracy, which boosts sensitivity by 9.1 per- cent.But this process is costly and time-consuming.Computer- aided diagnosis (CAD) is routinely used to identify and categorise breast masses. The use of computer- aided diagnosis increases precision and efficiency while reducing misclassifications. Computer-aided diagnosis, whichhas been demonstrated to be more effective than traditional methods, is frequently used to read mammograms a second time. To help doctors categorise breast cancer, the generic computer-aided diagnosis systemwhich includes phases for segmentation, feature extraction, and classification was cre- ated Therefore, an essential first step in creating computer- aided diagnosis systems is automatic picture segmentation. The general computer-aided diagnosis technique, which in- tegrates the processes of segmentation, feature extraction, and classification, was developed to assist medical professionals in classifying breast cancer. Therefore, an essential first step in the development of computer-aided diagnosis systems is automatic image segmentation. This review is divided into three sections: Section II discusses preprocessing techniques; Section III discusses segmentation and

   associated approaches. Finally, improved segmentation and preprocessing techniques.

2. PREPROCESSING

   Mammogram backgrounds can be removed using a variety of techniques so that the neural network is not disrupted by visual artefacts.However, none of those techniques make an effort to eliminate both the undesirable areas and the artefacts. Since adding some noise to the photos will make the train the data for DCNN appear more realistic, the neural networks performance will be enhanced for the pectoral muscle. In

   [1] conveys that training a neural network, adding noiseto the input data can improve generalisation performance. According to [2], adding Gaussian noise to the gradient during deep network training is beneficial. They discovered that the additional noise can aid in neural network model optimization. According to [3], the research used the rolling ball approach to minimise noise and find intensity level artefacts in the mammographic image. In order to spot micro-calcificationsin mammograms, the essential areas inside the breast were highlighted using the rolling ball algorithm. A ball with a specific radius is used in this process to roll over the images surface. It recognises a smooth, continuous background in the mammographic image. Based on the images intensity levels, the balls radius should at the very least be equal to the radius of the largest object. Construction of the ball, creation of a bright background, rolling of the ball, and removal of the background are all required. A genuine positive rate of 91.78 percent was reached.

   The mammograms are then processed to remove noise, artefacts, and superfluous areas using Huangs Fuzzy Thresh- olding Method and Morphological Transformations. After the rolling ball algorithm and background subtraction, the processed images that are stored on the local disc are trans- formed to binary images using the threshold values produced by Huangs approach [4]. The definition of membership func- tion and the idea of fuzzy sets are used in this image thresholding technique to assess the degree of fuzziness in an image and select the suitable threshold value. A method of image segmentation that reduces noise in photographs by using mathematical grey-scale morphology.Some of their methods characteristics, such as erosion and dilatation, have been incorporated into our suggested approach. They precisely located the breast borders and eliminated background from the images using morphological modifications known as erosion and dilatation. They were

   98.7 percent accurate overall.In [5] achieved an accuracy of 99.31 percent by using these morphological modifications (erosion and dilatation) to remove artefacts from their images. Theartefacts are then removed from the binarized image using morphological adjustments. Erosion and dilation have been used as transformation procedures in this study. The mam- mographic image is shrunk throughout the erosion process, making the brilliant spots smaller and the dark areas larger.On the deteriorated image, the dilation operation is carried

   out.The process makes the images bright parts larger. The bitwise AND operator is used to combine the rolling ball-processed image with the morphologically altered image in order to eliminate the artefact from the image. The backdropis eliminated after merging the two photos, and the resultsare then saved to a local drive. The Canny Edge Detection algorithm, created in 1986, is used to identify the pectoral muscles in the edge of the provided mammography picture [6]. The edges generated via deft edge detection have been detected using Hough Line Transform [7]. With the use of texture gradient and Euclidean distance regression, the muscle was distinguished from the mammogram by estimating the pectoral edge. The pectoral muscle could be eliminated from 96.75 percent of the photos by using this technique. For the purpose of segmentation and muscle boundary detection, it also utilised the Hough line transform. As shown by the methods 97.08 percent accuracy rate, the Hough transform is helpful for identifying the muscle boundary. After the relics have been eliminated, the muscle is the only thing that remains. Using Hough Line Transform and Canny Edge Detection, the muscle is removed. The excision of the muscle will be automated more quickly and easily if all of the mammographic images are smaller and facing the same way. The photos created by the merging process will therefore be initially flipped to the right side, then reduced in size. A convolutional neural network is utilised to detect the edges after the image hasbeen segmented using a method described by [8] basedon the pectoral muscle boundary. The disadvantage is that when employed in the real world, constructing and employing one CNN for pectoral muscle excision and then another CNN for detection and diagnosis can increase computation costs.

   Results of the suggested methods segmentation and detec- tion on a mammography image: (a) the original image, (b) the smoothed image, (c) the patch image following thresholding,

   1. the cancer region located in the input image in the window,

   2. the region patch located following morphological closing,

   3. the region boundary using gradient, (g) the cancer area detected, (h) the cancer area with region segmentation, and

      1. the proposed segmentation result of the cancer in the input mammogram image.

   After muscle has been eliminated using the Hough line transform, a Deep-CNN was developed for detection and diagnosis, and it may be used to analyse images with vis- ible muscle. In order to overcome the problem of curved boundaries,[9] this study employs image enhancing methods like Look-up Tables (LUTs) to assist the neural network in identifying and extracting ROIs and regions within ROIs. The idea of lookup tables for histogram equalisation was used byto equalise particular sections of chest pictures. According to, substantial information is lost when Xray film is digitalized using 12-bit quantization and then reduced to 8-bit pictures for display. They converted the photos into 4096 displayable pseudo colours to prevent this. The National Institutes of Healths ImageJ medical image processing software andthe Laboratory for Optical and Computational Instrumentss software, respectively, are used to implement the procedures. ImageJ is used to apply techniques for improving images, in- cluding Invert LUT, CTI RASLUT, and ISOCONTOUR LUT.Pre-

   processing of estimated results and the variations between those results with and without pre-processing. They managed to achieve a 95.42 percent accuracy without theuse of any pre-processing. Pre-processing made it possible to achieve 98.34 percent accuracy[10].

   Preprocessing is essential to enhancing the mammographic

   image quality and enhancing ROI extraction from the images. In [11] used colour coding to highlight the mammograms finer characteristics (pectoral muscle, fibro glandular tissue, breast tissue, backdrop, etc.), enabling the gathering of enough local information to categorise the pixels as belonging to various tissues and locations.

3. SEGMENTATION

   Prior to doing segmentation, pre-processing is crucial. U- Net based architecture is used for the segmentation of the tumour zone in histopathology images. The networks design has been enhanced and expanded to operate with less training photos and eliver more accurate segmentations. It is based on a fully convolutional network. The simultaneous use of global location and context is just one advantage of the U-Net technique for segmentation tasks. Second, even with less training examples, it performs better for segmentationtasks. The only glaring drawback of U-Net-style designs isthat network learning might neglect the layers where abstract properties are represented since learning might sluggish in the intermediate layers of deeper models.

   By AU-Net, the full mammograms may be processed. It also proposes a new up sampling block called the Attention Up (AU) Block and introduces asymmetrical structure to the usual encoder-decoder segmentation

   architecture. Three advantages in particular are intended to be provided by the AU block. Initial compensation for the information loss caused by bilin- ear up sampling is provided by dense up sampling. Second, it makes it easier to blend high-level and low-level features. It also provides a channel-attention feature in order to draw attention to channels with rich information. In [12] created a brand-new attention-guided dense-up sampling network for thorough mammography breast mass segmentation (AU-Net). The attention-guided dense-up sampling block and the efficient up sampling block are the two up sampling blocks that make up the asymmetrical encoderdecoder structure known as AU- Net (AU block). When tested using the two publicly accessible datasets CBIS-DDSM and INbreast, the approach yielded an average Dice similarity coefficient of

   81.80 percent for CBIS- DDSM and 79.10 percent for IN breast.

   The RU-Net model collects contextual information by fus- ing low-level and high-level components, and it shares the same basic architecture as the U-Net model. We redesignedthe U- Net topology by integrating residual attention modules to maintain spatial and contextual information, enable the networkto have a deeper architecture, and solve the gradient vanishing issue. The Res-Net classifier was used to classifythe data after

   [13] developed a residual deep learning strategy for a residual attention U-Net model (RU-Net) based on mass segmentation (RU-Net). The DDSM, BCDR-01, and IN breastdatasets were used to evaluate the suggested method, and noisewas decreased on all three datasets using the clare filter. The suggested model attained mean values of 94.0 percent for the mean test pixel accuracy,

   98.0 percent for the IOU, and 98.0 percent for the Dice coefficient index (DI). the innovative deep network By combining probabilistic graphical modelling with residual learning, Conditional Residual U-Net (CRU-Net) im- proves the performance of traditional U-Net segmentation. theinnovative deep network By combining probabilistic graphicalmodelling with residual learning, Conditional Residual UNet (CRU-Net) improves the performance of traditional U-Netsegmentation. By combining probabilistic graphical modelling and residual learning,conditional residual U-Net [14] enhancedbreast mass segmentation in mammograms (CRU-Net). The DDSM-BCRP and IN breast datasets, which are both openly accessible, were used to assess the CRU-Net technique. For theIN breast and the DDSMBCRP dataset, the CRU-Net attained Dice Index values of 93.66 percent and 93.32 percent re- spectively. For mass segmentation of digitised mammograms, deeply supervised UNet was developed (DS-U-Net). The clarefilter was used to improve the contrast in the images as the method was evaluated using the DDSM and IN breast datasets.Depending on whether the photos had been pre- processed or not, the tests were split into two groups. It was discoveredthat preprocessed trials produced findings that were superiorto those of unprocessed experiments. The preprocessing basedmethod produced 82.70 percent of Dice and 85.70 percentof Jaccard coefficients, 99.70

   percent accuracy, 83.10 percent sensitivity, and 99.80 percent specificity.

   The mixed-supervision-guided and residual aided classifi- cation U-Net model was created by [15] for the segmen- tation and classificationof mammography images (ResCU- Net).Convolutional filters were used to reduce the noise in the mammography images, which were acquired from the INbreast dataset. The suggested MS- ResCU-Net model outperformed ResCU-Net in all categories, scoring 94.16 percent accurately,

   93.11 percent sensitively, 95.02percent specifically, 91.78per- cent DI, 85.13percent Jac, and 87.22percent MCC. Using treereweighted belief propagation and deep learning potentials, mass mammography segmentation was suggested. The tech- nique was applied and evaluated using data from the INbreast and DDSM-BCRP databases, respectively, using a conditional random field model (CRF). The approach uses treereweighted belief propagation to reduce the mass segmentation error and statistical learning techniques to learn the data. The suggested method allowed for the quick and easy achievement of an 89.0 percent Dice index.

   A brand-new segmentation model for mammography pic- tures called the Fullresolution Convolutional Network (FrCN) was put forth by [16]. Using three popular deep learn- ing modelsconventional feedforward CNN, ResNet-50, and InceptionResNet-V2 the discovered and segmented breast tumours were also categorised as benign or malignant.The INbreast database was used to obtain the mammography images. FrCN had an overall accuracy of 92.97 percent for segmenting breast lesions, while MCC, Dice, and the Jaccard similarity coefficient had accuracy of 85.33 percent, 92.69percent, and 86.37 percent, respectively. In [17] they demonstrated that fully automatic breast density segmentation using deep learning and conditional generative adversarial networks (cGAN). A cGAN network was used to separate the thick tissues in mammography pictures. A median filter was used to reduce noise on 410 images of 115 patients from the INbreast dataset before the performance test. The accuracy, Dice coefficient, and Jaccard index values obtained from the cGAN segmentation were 98.0 percent

   , 88.0 percent, and 78.0 percent respectively.

   The adversarial deep structured net for mammography mass segmentation that is discussed is based on an end- to-end adversarial FCN-CRF network. The DDSM-BCRP and IN breast open datasets were used to evaluate the approach. The segmentation rate for the proposed method was 97.0 percent. For the identification, segmentation, and classification of breast masses, an integrated computerassisted diagnosis system based on deep learning and You-Only-Look-Once. It was recom- mended to partition the bulk using a local deep learning technique based on a full resolution convolutional network. When the technique was tested against the IN breast database, it generated a 99.24 percent F1 score,

   98.96 percent mass detection, and a 97.62 percent Matthews correlation coefficient (MCC).Using FrCN as the basis, mass segmentation accuracy was 92.97 percent

   , Jaccard similarity was 86.37 percent, MCC was 85.73

   percent , Dice was 92.69 percent, and so on. AUC was

   94.7 percent, MCC accuracy was 89.91 percent, and Dice accuracy was 968 percent. Additionally, the accuracy for mass detection and segmentation using CNN was 95.64 percent.

   Convolutional neural networks have been proposed by [18] for autonomous bulk segmentation in mammography physics. The semantic segmentation U-Net model, which was ini- tially created for biomedical picture segmentation applications,serves as the models foundation. Four databases were used forthe testing: CBIS-DDSM, IN breast, UCHCDM, and BCDR-Each database used an adaptive median filter as the noise- removal strategy. A mean Dice coefficient index and mean IOU of 95.10 percent and 90.90 percent, respectively, were generated by the suggested U-Net model. Because of the Dice coefficient index increases from 92.20percent to 95.10 percentand from 85.0 percent to 90.90 percent, respectively, the results are also improved by the use of data augmentation.

   In [19] it was proposed, that utilising cGAN and convolu- tional neural network to segment and classify breast tumours in mammograms. In a mammogram, a breast tumour is divided into a ROI by the cGAN. Performance was evaluated using DDSM data with 2620 mammography images and the INbreast dataset with 115 cases (a total of 410 mammograms). Noise was eliminated from the mammography pictures using morphological methods. It was suggested to use convolu- tional neural networks and cGAN. For the segmentation and categorization of breast tumours in mammograms. During a mammogram, the cGAN separates a breast tumour into a ROI. DDSM data with 2620 mammography images and the INbreast dataset with 115 patients were used to assess performance (a total of 410 mammograms).Noise was eliminated from the mammography pictures using morphological methods. In [20] it was able to segment and detect several breast lesions. A masked regional convolutional neural network with a feature pyramid network is the name of the method, which is based on a regional learning strategy. DDSM and the INbreast database were used for training and testing, respectively. The models segmentation accuracy was 91.0 percent , and the multi-detection mean average precision was 84.0 percent.To segment masses on mammography pictures using the U-Net algorithm, it is recommended using data augmentation. The DDSM databases 7989 mammography images were utilized to test the model. The model achieved a 79.39 percent Dice coefficient index, 92.32 percent sensitivity, 80.47 percent specificity, 85.95 percent accuracy, and 86.40 percent AUC. For the segmentation and classification of mammography pictures, [21] recommended a deep learning approach. Using the improved UNet model, the breast region was deleted from the mammography pictures. To test the model, three mam- mographic datasetsMIAS, DDSM, and CBISDDSMwere used. With respect to the DDSM datasets, the proposed model achieved a 97.99 percent F1 score, 98.87 percent accuracy, 98.88 percent area under the curve (AUC),

   98.98 percent sensitivity, and 98.76 percent precision.

4. CONCLUSION

The DNN gives better performance then ML segmentation techniques. In our review, it is observed that preprocessing is not required before training the Deep learning models. This may be the reasons for getting less accuracy in models without preprocessing. The present study helps to improve the accuracy and other quantification parameters by doing preprocessing initially and then implementing the training to the DNN model.

REFERENCES

[1] C. M. Bishop, Training with noise is equivalent to Tikhonov regularization, Neural Comput.,vol.7, no. 1, pp.108 116, Jan. 1995.

[2] A. Neelakantan, L. Vilnis, Q. V. Le, I. Sutskever, L. Kaiser,

K. Kurach, and J.Martens,Adding gradient noise improves learning for very deep networks, 2015, arXiv:1511.06807

[Online].Available: http://arxiv. org/abs/1511.06807.

[3] T. M. A. Basile, A. Fanizzi, L. Losurdo, R. Bellotti, U. Bottigli, R. Dentamaro, V. Didonna, A.Fausto,R. Massafra, M. Moschetta, P. Tamborra, S. Tangaro, and D. La Forgia, Mi- crocalcification detection in full-field digital mammograms: A fully automated computer-aided system, Phys.Medica, vol. 64,pp. 19, Aug. 2019.

[4] L.-K. Huang and M.-J.-J. Wang, Image thresholding by minimizing the measures of fuzziness,Pattern Recognit., vol. 28, no. 1, pp. 4151, Jan. 1995.

[5] D. A. Zebari, D. Q. Zeebaree, A. M. Abdulazeez, H. Haron, and H. N. A. Hamed, Improved threshold based and trainable fully automated segmentation for breast cancer boundary and pectoral muscle in mammogram images, IEEE Access, vol. 8, pp. 203097203116, 2020.

[6] J. Canny, A computational approach to edge detection, IEEE Trans. Pattern Anal.Mach.Intell.,vol. PAMI-8, no. 6, pp. 679698, Nov. 1986

[7] R. O. Duda and P. E. Hart, Use of the Hough transformation to detect lines and curves inpictures,Commun. ACM, vol. 15, no. 1, pp. 1115, Jan. 1972.

[8] H. Soleimani and O. V. Michailovich, On segmentationof pectoral muscle in digital mammograms by means of deep learning, IEEE Access, vol. 8, pp. 204173204182, 2020.

[9] Laboratory For Optical And Computational Instru- mentation. The University Of WisconsinMadison Research Lab Of Dr. Kevin Eliceiri. Accessed: Dec. 7, 2020. [On- line].Available:https://www.eliceirilab.org

[10] ImageJ Software Tool. Accessed: Dec. 7, 2020. [On- line].

Available: https://www.imagej.net/

[11] A. Dubrovina, P. Kisilev, B. Ginsburg, S.Hashoul, and

R. Kimmel, Computational mammography using deep neural networks, Comput. Methods Biomech. Biomed. Eng., Imag. Vis.,vol. 6, no. 3, pp.243247, 2018.

[12] H. Sun, L. Cheng, B. Liu et al., Aunet: attention- guided dense-upsampling networks for breast mass segmen- tation in whole mammograms, Physics in Medicine amp;Biology, vol. 65, no. 5, p.55005,2020.

[13] D. Abdelhafiz, S. Nabavi, R. Ammar, C. Yang, and J. Bi, Residual deep learning system for mass segmentation and classification in mammography, in Proceedings of the 10th ACM International Conference on Bioinformatics, Computa- tional Biology and Health Informatics, pp. 475 484, Niagara Falls, NY, USA,2019

[14] H.Li, D. Chen, W. H. Nailon, M. E. Davies, and D. Laurenson, Improved breast mass segmentation in mam- mograms with conditional residual u-net, in Image Analysis for Moving Organ, Breast, and Thoracic Images, pp. 8189, Springer, Cham, 2018.

[15] T. Shen, C. Gou, J. Wang, and F. Y. Wang, Simul- taneous segmentation and classification of mass region from

mammograms using a mixed-supervision guided deep model, IEEE Signal Processing Letters, vol. 27, pp. 196200, 2019.

M. A. Al-antari, M. A. Al-masni, and T. S. Kim,

[16] DeeLearning Computer-Aided Diagnosis for Breast Lesion in Digital Mammogram, in Advances in Experimental Medicineand Biology, pp. 5972, Springer, Cham, 2020.

[17] N. Saffari, H. A. Rashwan, M. Abdel-Nasser et al., Fully automated breast density segmentation and classifica- tion using deep learning, Diagnostics, vol. 10, no. 11, p.

988, 2020.

[18] D. Abdelhafiz, J. Bi, R. Ammar, C. Yang, and S. Nabavi,Convolutional neural network for automated mass segmentation in mammography, BMC Bioinformatics, vol. 21, no. S1, pp. 119,2020

[19] V. K. Singh, H. A. Rashwan, S. Romani et al., Breast tumor segmentation and shape classification in mammograms using generative adversarial and convolutional neural net- work,Expert Systems with Applications, vol. 139, p. 112855,2020

[20] H. M. A. Bhatti, J. Li, S. Siddeeq, A. Rehman, and Manzoor, Multi-detection and segmentation of breast lesions based on mask rcnn-fpn, in 2020 IEEE International Conference on Bioinformatics and Biomedicine (BIBM), pp. 26982704, Seoul, Korea (South), 2020.

[21] W. M. Salama and M. H. Aly, Deep learning in mam- mography images segmentation and classification: automated cnn approach, Alexandria Engineering Journal, vol. 60, no.5, pp.47014709, 2021.