Computational methods using machine learning for identification and automatic classification of mammary nodules and microcalcifications in digital mammography exams

Abstract

Statistical data show that around 1.67 million new cases of breast cancer are diagnosed annually worldwide, and that about 500,000 patients die each year because of this pathology. In Brazil, about 60 thousand cases were registered in the last year, with a mortality rate of 25%. Breast cancer consists of the development and evolution of carcinomas in the breast tissue, formed mainly by the evolution of nodules, microcalcifications and other lesions, and the early diagnosis is fundamental to increase the patient's chances of recovery. Mammography is considered the best method of diagnosis, capable of detecting these changes in breast tissue, the only one able to detect breast cancer prematurely, and fundamental for the BIRADS risk classification. However, diagnosis only through the visual evaluation of these exams, without any digital processing, may lack information for a quick and accurate analysis. Recent studies present computational methods that allow the identification and classification of these diseases, aiming to overcome these visual limitations, but still have limitations in terms of accuracy and comprehensiveness. Therefore, the proposed project aims to investigate and evaluate computational approaches of computer classification combining techniques of image processing, extraction of characteristics and machine learning using images of mammography exams. These approaches will be developed and evaluated, and tools will be developed to automatic identify and classify breast abnormalities and lesions. It is intended to perform the identification, segmentation and classification of breast lesions such as nodules, cysts and microcalcifications through neural networks and machine learning. For this, two approaches will be evaluated, one using deep learning through the Convolutional Neural Network (CNN), and the other using conventional neural networks performing "attribute engineering". Hoping to result in a methodology with greater speed, precision and accuracy in locating, identifying and monitoring these lesions, and embarking on an online platform to facilitate access and use by radiologists. This platform will connect our tools to online PACS systems, which will make use of solutions available for hospitals, clinics, imaging laboratories, radiologists and teleradiology companies. Performing a pre-diagnosis from the analyzes of specific neural networks, and issuing a report with notes and classifications for the radiologist. In order to make it more practical and accessible to use tools to aid the diagnosis of mammography examinations, and to increase the accuracy and agility in the issuance of mammography examination reports. Allowing patients to initiate appropriate treatment as soon as possible, and increasing their chances of recovery. And expecting to reach about 5% of this national marking in the first two years of operation (about 500 thousand examinations) and providing a turnover of about 3 million of reais for the company only with mammography exams in the national market. (AU)