Room: AAPM ePoster Library
Purpose: Accurately segmenting tumors plays an important role in the diagnosis and treatment of breast cancer patients. We aim to develop a novel fully automatic tumor segmentation model for digital breast tomosynthesis (DBT) images through U-net.
Methods: A total of 50 DBT images are used in this study. Tumors manually contoured by experienced oncologists are considered as the ground truth. The model was designed based on U-net architecture. 25 images were randomly selected as the training dataset and data augmentation was performed to generate 1,500 new images for training. All these images are fed into a new designed U-net architecture to train the segmentation model. The testing stage consists of three steps: (1) tumor segmentation through trained U-net; (2) the minimum thresholding method was applied for the post-processing; and (3) segmentation results were evaluated using true positive rate (TPR), false positive rate (FPR) and F-Score.
Results: The average results of quantitative analysis for the TPR, FPR, and F-Score are 0. 889, 0.002, and 0.918, respectively.
Conclusion: We developed an automatic tumor segmentation strategy for DBT image, and segmentation performance is promising to consider no human assistance or interaction.