Resumen de Analyzing the breast tissue in mammograms using deep learning
Breast density is a prevalent high-risk cause of breast cancer in women. Numerous studies have reported that women with high-density breasts develop breast cancer independent of their age, menopausal status, and hormone replacement therapy.
Therefore, improving the screening process for women with high breast density is vital and can decrease the risk of breast cancer, so it is essential to report the breast density to guide women with high breast density for additional examinations, such as ultrasonography or MRI.
The main goal of this doctoral dissertation project was to investigate the fully automated segmentation and classification scheme for mammograms based on breast density estimation using deep learning methods to overcome this challenge. The findings presented in this doctoral dissertation is promising and show that the proposed technique can produce a clinically helpful computer-aided digital tool for breast density analysis by digital mammography.
The present thesis makes several noteworthy contributions to the literature on developing innovative algorithms based on machine learning and deep learning techniques to improve and increase the accuracy of fully automated mammograms computer-aided design (CAD). In addition, this research has several practical applications, which are presented in four chapters and summarized below.
Chapter 2 proposes a new methodology by combining the traditional texture analysis methods with deep learning approaches to classify breast tumors into malignant or benign. This method used a bio-inspired optimization algorithm (also known as a whale optimization algorithm) to learn local descriptors from the input images. Ultrasound and mammographic images datasets with benign and malignant cases were used to assess the proposed method. This research showed the accuracy of this model on both ultrasound and mammographic images was within acceptable ranges.
In Chapter 3, a novel breast density segmentation methodology is developed based on the conditional Generative Adversarial Network. The cGAN network consists of two networks, the first network is an encoder-decoder network (also called the generator) that maps the input mammogram to a segmented image including the dense areas, and the second network is called discriminator that learns a loss function to enforce the generator to produce an output, which is like the ground-truth. Using the cGAN network helps the proposed model to train by a small set of mammographic images. Furthermore, the proposed method can segment the dense regions in mammographic images with outstanding performance compared to the standard semantic segmentation methods (FCN and SegNet).
In Chapter 4, an innovative and accurate methodology is developed for the segmentation and classification of breast density based on the BI-rads standard. Unfortunately, traditional breast density segmentation and classification methods are complex tasks and have a high possibility of false positives. To overcome this problem, the efficacy of a fully automated algorithm for breast density segmentation and classification in digital mammography is proposed and substantiated by presenting three versions of cGAN networks for egmentation and two different classification methods.
In the experiments carried out in this chapter, mammograms of 115 patients (410 images) from the INbreast dataset were utilized. The breast density segmentation study showed that the proposed cGAN with dice loss function could achieve achieved an accuracy of 98%, which is the highest accuracy compared to other loss functions.
In addition, using the proposed CNN methodology for density classification leads to an accuracy equal to 98.75%. Finally, it is noteworthy that a strong correlation between the computerized algorithm’s output and the radiologist’s estimated breast density can be obtained. This observation justifies that the proposed methods in this chapter have a strong positive relationship with the radiologist manual classification and are competitive with reported correlation coefficients from the literature.
For future developments, more datasets need to be utilized; however, the ground-truth of the dataset needs to be prepared by doctors, experts, and radiologists of the Hospital Universitari Sant-Joan of Reus, Spain, via developing our Graphical User Interface (GUI) in MATLAB to help the radiologists to annotate the images. It is believed that artificial intelligence is capable of surpassing human experts in breast density prediction.
One of the most important practical applications of this Ph.D. project which can emerge in the near future, will be transposing the fully automated PD% estimation techniques developed in this thesis into a robust computer-aided breast density analyzer appraisal tool/software for the public use in clinical practice.
Chapter 5 analyzes the local and general breast density changes in temporal follow-up mammograms after treatment based on deep learning methods. For tissue segmentation, the cGAN technique was applied to learn the intrinsic features of the density from a relatively low number (hundreds) of training samples and then generate the corresponding image mask that selects the pixels belonging to the tissue.
The main aim of this methodology is to determine the threshold between risky and non-risky cases. If the global and local breast density difference between control “n” and “n-1” of patient” i” is more than a threshold, patient” i” is considered as a risky case and requires more analysis.
Overall, this thesis presents the automated deep learning method for breast density analysis; This thesis 's findings are promising and show that the proposed techniques can produce a clinically helpful computer-aided tool for breast density analysis by digital mammography.
