Breast density prediction from simulated mammograms using deep learning

High breast density (BD) is recognized as an independent risk factor for breast cancer development, in addition to negatively impacting the sensitivity of mammography by hiding tumor masses. Although BD is normally assessed with the BI-RADS reporting system, this evaluation is qualitative and has been shown to vary considerably across readers, which usually divide density into four different classes. In this study, it’s presented a deep learning (DL) method to quantify BD from a standard two-view (cranio-caudal, and medio-lateral-oblique) mammography exam. With the aim of developing a method based on an objective ground truth, the DL model was trained and validated using 88 simulated mammograms from an equal number of distinct 3D digital breast phantoms for which BD is known. The phantoms had been previously generated through segmentation and simulated mechanical compression of patient dedicated breast CT images, allowing for the exact calculation of BD in each case. Different augmentations were applied prior to simulation, to increase the dataset size and take into account the variability among women. These augmentations included different breast size and different proportion between the two main breast tissues (fibroglandular and adipose) and have led to a total of 528 cases. These were divided, randomly and on a patient level, into training (N=360), validation (N=60), and test sets (N=108), making each set adequately represents the density scale (from 0 to 100 in percentage density). Considering the shape assumed by the breast during the mammography examination, an additional DL model (U-Net) has been implemented for the segmentation of the zone with constant thickness directly from mammograms, with the aim of finding the only adipose pixel for which to standardize the values. This operation, initially performed manually, required a U-Net to exclude the use of tissues thickness maps (which would not be available in the clinic) and thus make the algorithm fully automatic. The DL prediction model performance was tested by stratifying the breasts into four different density ranges: 5-15%, 15-25%, 25-60%, and >60%. The median absolute errors and interquartile ranges (IQR), in percentage points, were: 3.3 (IQR: 3.5), 3.4 (IQR: 2.5), 3.5 (IQR: 3.9), and 14.8 (IQR: 8.4), respectively. These results were obtained by applying the model on the test set’s mammographies from the same vendor (Siemens Mammomat Inspiration). However, the model seems to accurately predict the density also when applied on other system’s images, without the need of re-training. When tried on a different vendor (Hologic Selenia Dimension), indeed, the median absolute errors and the interquartile ranges in the same ranges as before were: 4.5 (IQR: 4.3), 3.5 (IQR: 3.6), 3.4 (IQR: 5.1), and 24.2 (IQR: 17.6), respectively. Although preliminary, these results show the potential of the proposed approach for accurate BD quantification, which is based, as opposed to most previously proposed approaches, on an objective ground truth.