Automated Assessment of Breast Positioning Quality in Screening Mammography.

Abstract

Simple SummaryInadequate breast positioning quality is the main cause behind misdiagnosis of breast cancer in screening mammography. For this reason, the first step before any cancer diagnosis is to ensure that the acquired mammograms have adequate breast positioning quality according to predefined criteria. If this is not the case, the patient must return for a new mammography. In this study, we proposed an approach for an automatic assessment of breast positioning quality in screening mammography using Convolutional Neural Networks. Eventually, this approach is not intended to replace radiology technicians, but rather to assist them in identifying inadequately positioned mammograms in real time, reduce the number of returned patient and improve the efficiency of cancer detection. For each predefined criterion, a specific convolutional neural network was separately trained and then combined into an overall system that predicts whether the breast is well positioned or not, achieving an efficient accuracy of 96.5% for craniocaudal and 93.3% for mediolateral oblique images. Our approach differs from already available studies and commercial tools by taking into account more useful breast positioning criteria that have to be considered by the expert, and thus providing a more holistic assistance.Screening mammography is a widely used approach for early breast cancer detection, effectively increasing the survival rate of affected patients. According to the Food and Drug Administration’s Mammography Quality Standards Act and Program statistics, approximately 39 million mammography procedures are performed in the United States each year. Therefore, breast cancer screening is among the most common radiological tasks. Interpretation of screening mammograms by a specialist radiologist includes primarily the review of breast positioning quality, which is a key factor affecting the sensitivity of mammography and thus the diagnostic performance. Each mammogram with inadequate positioning may lead to a missed cancer or, in case of false positive signal interpretation, to follow-up activities, increased emotional burden and potential over-therapy and must be repeated, requiring the return of the patient. In this study, we have developed deep convolutional neuronal networks to differentiate mammograms with inadequate breast positioning from the adequate ones. The aim of the proposed automated positioning quality evaluation is to assist radiology technologists in detecting poorly positioned mammograms during patient visits, improve mammography performance, and decrease the recall rate. The implemented models have achieved 96.5% accuracy in cranio-caudal view classification and 93.3% accuracy in mediolateral oblique view regarding breast positioning quality. In addition to these results, we developed a software module that allows the study to be applied in practice by presenting the implemented model predictions and informing the technologist about the missing quality criteria.