Automated Breast Lesion Segmentation from Ultrasound Images based on PPU-Net

Ultrasound is one of the most widely applied imaging modalities for breast lesion assessment. Breast lesion segmentation in ultrasound plays an important role in extracting features that is crucial for breast cancer diagnosis and treatment planning. Existing methods include manual or automated identifying breast lesion boundary. However, manual segmentation is time consuming and leads to inter- and intra-observer variations. Automated segmentation of breast lesion is highly required but very challenging due to the ultrasonic property of tissue. The neural network, an emerging automated segmentation technology, is widely used in medical images segmentation tasks and has achieved good performance. In this study, we proposed a novel pyramidal pooling U-Net network (PPU-Net) to segment breast lesion. In PPU-Net, the pyramid pooling module (PPM) was applied together with the U-Net to extract more scale information. The ratio of training set (408 cases) and test set (103 cases) was 4:1 in 511 breast lesion images. The PPU-Net approach was evaluated on the test set and achieved good performance. The evaluation metrics of Dice similarity coefficient (DSC) and accuracy (ACC) for PPU-Net were $88.97 \pm 8.96$% and $94.16\pm 4.02$% respectively. For comparision, the DSC and the ACC based on U-Net and FCN were evaluated. The result demonstrated the feasibility that PPU-Net could be applied in breast lesion segmentation from ultrasound images and it has better performance than U-Net and FCN in this study.