Towards accurate breast tumor segmentation in ultrasound images using center representation learning
Guanqun Ding, Kaiwen Yang, Ziyi Chen, Zizhi Huang · 2025
Accurate segmentation of breast tumors in ultrasound images is essential for early diagnosis and treatment planning in breast cancer care. However, challenges such as low contrast and indistinct tumor boundaries often result in misaligned tumor center predictions, which hinders precise tumor identification. In this study, we propose a novel segmentation model that improves accuracy by learning tumor center representations from ultrasound images. The model utilizes an encoder-decoder architecture, where the encoder extracts multiscale visual features. The decoder consists of two branches: a segmentation branch that progressively fuses multi-scale features to predict the tumor segmentation map, and a center embedding branch that integrates multi-scale center-related embeddings to generate a tumor center map. To guide the segmentation branch in locating the tumor center, we introduce a tumor center guidance module that merges multi-level tumor center embeddings within the decoder. Extensive experiments on several public breast tumor segmentation datasets demonstrate that our method outperforms existing approaches, validating its effectiveness in improving segmentation accuracy.