Developing Large Pre-trained Model for Breast Tumor Segmentation from Ultrasound Images

Early detection and diagnosis of breast cancer using ultrasound images are crucial for timely diagnostic decision and treatment in clinical application. However, the similarity between tumors and background and also severe shadow noises in ultrasound images make accurate segmentation of breast tumor challenging. In this paper, we propose a large pre-trained model for breast tumor segmentation, with robust performance when applied to new datasets. Specifically, our model is built upon UNet backbone with deep supervision for each stage of the decoder. Besides using Dice score, we also design discriminator-based loss on each stage of the decoder to penalize the distribution dissimilarity from multi-scales. Our proposed model is validated on a large clinical dataset with more than 10000 cases, and shows significant improvement than other representative models. Besides, we apply our large pretrained model to two public datasets without fine tuning, and obtain extremely good results. This indicates great generalizability of our large pre-trained model, as well as robustness to multi-site data. The code is publicly available at https://github.com/limy-ulab/US-SEG .