Artificial Intelligence System for Automatic Mammary Region Extraction Using Semi-subjective Corrected Region for Breast Composition Evaluation

Introduction Recently, breast composition has been used as a clinical indicator for breast cancer. Although systems have been developed for objectively extracting mammary gland regions, relying on subjective judgment to identify correct mammary gland regions for breast composition can lead to significant inter-judgment variation. In this study, we automatically extracted mammary gland regions using semi-subjective corrected regions that extract only mammary gland regions while simultaneously determining quantitative regions and examining whether extracted results could be used clinically. Methods We used 670 mammograms (Pe-ru-ru, Canon Medical Systems Corporation, Tochigi, Japan). A breast physician with 30 years of experience reading mammograms subjectively evaluated mammary gland regions based on the quantitatively determined regions. We defined these images as semi-subjective corrected region images. Further, we used U-Net for segmentation and the dice coefficient as the evaluation index for the region extraction accuracy. The parameters of U-Net (number of downsampling layers, learning rate, and batch size) and the orientation of input images were changed to improve accuracy. In addition, we calculated the dice coefficient based on the breast composition type to evaluate the clinical usefulness of this study. Results The average dice coefficient with the highest accuracy was 0.882; the average dice coefficients were 0.992, 0.832, 0.904, and 0.943 for fatty, scattered, heterogeneous dense, and extremely dense regions, respectively. Conclusion The mammary gland region was automatically extracted using semi-subjective corrected region images. The average dice coefficients for the whole breast and for each breast composition were high, suggesting that this method is clinically useful.