Breast density in dedicated breast computed tomography

Abstract

The aim of this study was to develop a new breast density classification system for dedicated breast computed tomography (BCT) based on lesion detectability analogous to the ACR BI-RADS breast density scale for mammography, and to evaluate its interrater reliability. In this retrospective study, 1454 BCT examinations without contrast media were screened for suitability. Excluding datasets without additional ultrasound and exams without any detected lesions resulted in 114 BCT examinations. Based on lesion detectability, an atlas-based BCT density (BCTD) classification system of breast parenchyma was defined using 4 categories. Interrater reliability was examined in 40 BCT datasets between 3 experienced radiologists. Among the included lesions were 63 cysts (55%), 18 fibroadenomas (16%), 7 lesions of fatty necrosis (6%), and 6 breast cancers (5%) with a median diameter of 11 mm. X-ray absorption was identical between lesions and breast tissue; therefore, the lack of fatty septae was identified as the most important criteria for the presence of lesions in glandular tissue. Applying a lesion diameter of 10 mm as desired cut-off for the recommendation of an additional ultrasound, an atlas of 4 BCTD categories was defined resulting in a distribution of 17.5% for density A, 39.5% (B), 31.6% (C), and 11.4% (D) with an intraclass correlation coefficient (ICC) among 3 readers of 0.85 to 0.87. We propose a dedicated atlas-based BCTD classification system, which is calibrated to lesion detectability. The new classification system exhibits a high interrater reliability and may be used for the decision whether additional ultrasound is recommended.

Applied Machine Learning in Spiral Breast-CT: Can We Train a Deep Convolutional Neural Network for Automatic, Standardized and Observer Independent Classification of Breast Density?

“…Therefore, a system for standardized reporting of BCTD as a surrogate for the sensitivity of the examination analogous to the ACR BI-RADS scale is greatly needed. Wieler et al [ 18 ] proposed a four-level classification system (A to D) intended for human reading with the relevant features amount of breast parenchyma, distribution of breast parenchyma and visibility of fatty septae between glandular tissue, distinguishing the different classes. Earlier studies by Stomper et al regarding the breast density distribution in the female population in conventional mammography reported similar distribution as in our patient cohort.…”

“…In our study, we propose an approach to determine breast density, according to Wieler et al [ 18 ], using a deep convolutional neural network (dCNN), which is the most powerful machine learning algorithm for the classification of radiological images, however, it also requires a very large amount of data to reach satisfactory accuracy levels. Our four evaluated dCNN configurations were trained with approximately 6000 images of spiral breast-CT, which were labelled according to breast density by human reading.…”

“…Human reading was performed by a radiology resident with one year of experience in mammographic imaging, regarding raw data images in the coronal plane. BCT examinations were categorized in four groups, with different breast densities based on lesion detectability, according to Wieler et al [ 18 ], as depicted in Figure 1 .…”

“…: 16 years of experience). Both readers were blinded to patient information and rated each image of the BCT examination individually, according to the provided density atlas by Wieler et al [ 18 ] The difference in image evaluation in our study compared to Wieler et al was that only raw images in the coronal plane were used to train the dCNN and for the human read-out. The classification of the dataset of both readers served as a standard for the evaluation of the classification accuracy of the dCNN and was used for computation of inter-reader agreement between both readers.…”

“…Yet, the BI-RADS atlas has been developed for conventional projection mammography and does not consider the three-dimensional nature of BCT imaging datasets. A new dedicated BCT breast density classification atlas has recently been proposed by Wieler et al [ 18 ], implementing fatty septae into the breast density assessment, which are unique in BCT and allow for the improved distinction of lesions within parenchyma. In our study, breast density was assessed by human reading, using the provided atlas by Wieler et al and we subsequently trained a deep convolutional neural network (dCNN) for automatic breast CT density (BCTD) classification.…”