Assessment Of Mammogram Images For The Prediction Of Breast Cancer Using Machine Learning Techniques

Breast cancer is the 2nd leading cancer of death among women around the world.In Asia and Africa,due to low income,the mortality rates are very high compared to Europe and America.According to the World Health Organization International Agency for Research on cancer(WHO-IARC)report 2020,2030000 new cases were diagnosed with breast cancer among women worldwide,and 11.6% among all other cancer and mortality rates in the same year was 685000,which is 6.6%.Breast cancer begins when the tumor cell grows out of control.These tumors can be recognized on an X-ray machine.Mammogram images are the standard screening method to detect breast cancer at early stages.Automatic detection of the lesion in mammograms with computer algorithms helps radiologists to detect breast cancer easily in less time and forbid unnecessary biopsies.Computer aided diagnostic(CAD)system mainly focuses on to remove noise from the images and to detect abnormalities which further assist radiologist for the better prediction of cancerous regions.These tumors are malignant if the cell invades and spread into another area of the body.Most malignant tumors start in milk ducts.It is observed that various image processing modalities have been used to discriminate breast density,architectural distortion and microcalcification in mammograms to detect breast cancer.Microcalcification(MC)and architectural distortion(AD)are the primary abnormalities in human breast for the early signs of breast cancer.Initially,image interpretation is manually conducted by the radiologist and physicians that requires expertise;thus,the CAD system is necessary to enhance the accuracy of cancer diagnostics in mammograms at early stages.Several feature extractions,pattern recognition and classification techniques have been developed by the programmers and published by the researcher in academic journals.Researchers have proposed several CAD systems in the past few years to predict breast cancer from mammogram images.However,no such method is available for assessing mammogram images into grades for detecting breast cancer.Furthermore,most researchers utilized traditional hand-crafted machine learning techniques for feature extraction and object detection,which is time-intensive and relies on professional radiologists.These studies were evaluated on small and unbalanced data,which led to unreliable results.This dissertation proposed several novel methods for automatically detecting of breast cancer from digital mammograms based on image assessment.A concrete framework is proposed,highlighting the importance of image preprocessing techniques.The proposed study deals with data imbalanced and feature extraction with an appropriate predictive model,which can help radiologists and researchers to diagnose breast cancer in its early stages.This research aims to counter the above limitation and to develop a CAD system to the assessment of mammogram images into benign and malignant grades by achieving higher accuracies and true positive rates.Following are the main contributions of this dissertation.1.Firstly,this dissertation examines the challenge of individual MC cluster detection in suspicious regions by proposing an automated de-noising method using Gaussian notch-based detector and wavelet transformation with depth-wise separable convolutional neural network.The concrete image pre-processing and segmentation techniques,such as image resizing,fetching suspicious and non-suspicious pixel values,handling missing values,and noise reduction,using a new Gaussian notch-based detector was,employed to optimize classifiers’ performance and filter out most deterministic suspicious factors associated with them.Furthermore,equalize the mammogram noise by calculating the noise statistics using a proposed Gaussian notch-based detector to improve the contrast enhancement of microcalcification detection.The calcification segmentation is performed using two-dimensional wavelet transformation and wavelet coefficients subbands LL,LH,HL,and HH filter bank for feature extraction.Finally,wavelet-based,fully connected,depth-wise,separable CNN architecture was developed for classifying these segmented MC clusters into malignant and benign grades.The presented model performs well in detecting and classifying MCs for the prediction of breast cancer.The presented model performs well in detecting and classifying MCs for the prediction of breast cancer.2.Secondly,this dissertation investigates the problem of tracking the cancerous and non-cancerous AD region of interests(ROIs)on mammography projection by employing a novel approach to facilitate radiologists by improving the diagnostic ability of the deep learning model.An adversarial gradient guided VGG-19 deep convolutional neural network was developed for the classification of AD ROIs to predict breast cancer.A two-dimensional contourlet transform was employed for segmentation of suspicious AD ROIs using low and high frequency coefficients.The edge detail information of AD and noise was evident in the high-frequency sub-bands after decomposing the mammography with contourlet transform.The adversarial gradient guided loss function was employed to preserve AD ROIs and adversarial loss to improve the performance of VGG-19 deep convolutional neural network.We trained the proposed deep convolutional neural network using a weighted combination of mean squared error loss and adversarial loss.The presented deep learning model outperforms in classifying the AD ROIs and improves the cancer prediction accuracy.3.Finally,this study counters the limitation of texture-based AD tracks detection by reducing the actual false cases and increasing the rate of truly diagnosed cases without the help of experts using an automated model.A novel image fusion framework was proposed using a three-dimensional depth-wise VGG-16 convolutional neural network for image-wise feature extraction and classification of AD ROIs to predict breast cancer.The proposed method pertains to the steps such as image preprocessing,enhancement,texture generation and segmentation using computer vision algorithm.The feature fusion filter bank was created using a Laplacian of Gaussian for the classification of these ROIs with three-dimensional depth-wise VGG-16 convolutional neural network.Our proposed method achieving a higher true positive rate as compared with previous approaches.This study establishes a fundamental theory for developing an automated computer diagnostic system to predict,assist radiologists,health professionals,and alleviate breast cancer prevalence.This research could be an interesting and valuable study in the healthcare predictive modeling domain and will make a real contribution to society.