Multi-Objective Representation based Dynamic Prototype Learning for Unsupervised DCE-MRI Breast Tumor Segmentation

Abstract

Unsupervised segmentation is a potential means to detect the breast tumors from DCE-MRI without using any annotated images, which can provide a coarse prior for several downstream tasks. However, the existing unsupervised segmentation methods are prone to collapse due to the presence of large background regions in breast DCE-MRI and the absence of effective constraints. To address these issues, we proposed a dynamic prototype learning network (DyPNet) based on nearest moving average (NMA) strategy and multi-cluster reconstruction (MCR) constraints for segmenting breast tumors from DCE-MRI unsupervisedly. Specifically, MCR is used to constraint the feature embeddings, ensuring the completeness and specificity of features for each cluster, and NMA is used to adaptively update the cluster prototypes according to the intensity of top-L intra-cluster samples. Using the Cosine similarity between the prototypes and the pixel embedding features, the image pixels can be clustered. By comparing the proposed method with several unsupervised segmentation models on different datasets, we demonstrated that the averaged DSC, HD95, and VM of the proposed method can be improved by 73.7%, 20.7% and 111.3% respectively. In addition, through the comparisons on downstream tasks, including one-shot, generalized and zero-shot segmentations, we further verified the effectiveness and superiority of the proposed method.