Automated fibroglandular tissue segmentation in breast MRI using generative adversarial networks.

摘要: Fibroglandular tissue (FGT) segmentation is a crucial step for quantitative analysis of background parenchymal enhancement (BPE) in magnetic resonance imaging (MRI), which useful breast cancer risk assessment. In this study, we develop an automated deep learning method based on generative adversarial network (GAN) to identify the FGT region MRI volumes and evaluate its impact specific clinical application. The GAN consists improved U-Net as generator generate candidate areas patch convolutional neural (DCNN) discriminator authenticity synthetic region. proposed has two improvements compared classical U-Net: (1) designed extract more features accurate description region; (2) DCNN discriminating generated by U-Net, makes result stable accurate. A dataset 100 three-dimensional (3D) bilateral scans from patients (aged 22-78 years) was used study with Institutional Review Board (IRB) approval. 3D hand-segmented all breasts were provided reference standard. Five-fold cross-validation training testing models. Dice similarity coefficient (DSC) Jaccard index (JI) values evaluated measure accuracy. previous using baseline study. five partitions set, achieved DSC JI 87.0 ± 7.0% 77.6 10.1%, respectively, while corresponding obtained through 81.1 8.7% 69.0 11.3%, respectively. significantly superior U-Net. impacted BPE quantification application following manner: correlation coefficients between quantified value BI-RADS categories radiologist 0.46 0.15 (best: 0.63) segmented areas, 0.41 0.16 0.60) areas. can be better model than