Semi-BGSEGNET: A Semi-Supervised Boundary-Guided Breast Tumor Segmentation Network

Segmentation allows localization and visualization of breast lesions of interest, thus essential for precise diagnosis, prognosis and treatment. However, accurate segmentation from magnetic resonance imaging (MRI) faces following challenges: 1) multiscale breast tumors with a large size range; 2) image quality degradation caused by motion artifacts; 3) blurred tumor boundaries especially malignant tumors; 4) limited labeled MR images. To address these challenges, we propose a semi-supervised boundary-guided segmentation network (Semi-BGSegNet) for breast tumors. Specifically, we first construct a boundary-guided baseline segmentation network (BGSegNet) that integrated multi-scale and attention modules to the encoder-decoder architecture. A dynamic boundary loss is combined with the cross-entropy loss to guide the update of BGSegNet. Then, we develop a lightweight pixel-level discriminator to discriminate predictions from labels. The discriminator can also provide pseudo labels for unlabeled data for semi-supervised training of BGSegNet. Experiments demonstrated the effectiveness of our model compared to other supervised and semi-supervised methods.