Background: In the field of breast screening using mammography, announcing to the examinees whether they are dense or not has not been deprecated in Japan. One of the reasons is a shortage of objectivity estimating their dense breast. Our aim is to build a system with deep learning algorithm to calculate and quantify objective breast density automatically.
Material and method: Mammography images taken in our institute that were diagnosed as category 1 were collected. Each processed image was transformed into eight-bit grayscale, with the size of 2294 pixels by 1914 pixels. The "base pixel value" was calculated from the fatty area within the breast for each image. The "relative density" was calculated by dividing each pixel value by the base pixel value. Semantic segmentation algorithm was used to automatically segment the area of breast tissue within the mammography image, which was resized to 144 pixels by 120 pixels. By aggregating the relative density within the breast tissue area, the "breast density" was obtained automatically.
Result: From each but one mammography image, the breast density was successfully calculated automatically. By defining a dense breast as the breast density being greater than or equal to 30%, the evaluation of the dense breast was consistent with that by a computer and human (76.6%).
Conclusion: Deep learning provides an excellent estimation of quantification of breast density. This system could contribute to improve the efficiency of mammography screening system.
Keywords: breast screening; deep learning; dense breast; semantic segmentation.