Breast-Density Semantic Segmentation with Probability Scaling for BI-RADS Assessment using DeepLabV3

Mammographic breast density is an early indicator of a patient's risk for breast cancer development. Although the direct cause is not fully understood, increased mammographic breast density increases the chance of developing breast cancer. Based on the density pattern exhibited, a patient's breast density score is assigned by a radiologist using one of four categories outlined by the Breast Imaging Reporting and Data Systems (BI-RADS). The subjective nature of pattern-based assessment of patient breast density can result in increased intra- and inter-rater variability. The identification of objective quantitative scoring systems for breast density lends the opportunity to provide patients and radiologists with increased granularity and precision over the four subjective BI-RADS classes. We propose a quantitative breast-density approach using semantic segmentation to assign each of a patient's breasts with a density value. Using the values for each breast, a patient-level density value can be used to determine the patient's placement on a linear density scale along with the probability that the patient belongs to each of the BI-RADS breast-density classes. We use accuracy and kappa scores to determine the performance of the algorithm with BI-RADS classification. Results show comparative classification and performance can be achieved between our semantic segmentation methodology and previous work which implemented a Vision Transformer (ViT)-based breast-density classification algorithm.