DAUS-Net: Toward Ultrasound Scanner-Agnostic Domain Generalized Robust and Accurate Segmentation

Abstract

In medical imaging, segmentation is a critical task for analysis and diagnosis. Deep learning-based segmentation has been actively studied and has shown remarkable performance. Building high-accuracy segmentation models requires a large amount of high-quality labeled data, but the cost of collecting such data is extremely high in medical imaging. In ultrasound imaging, the differences in image features depending on the equipment are significantly greater compared to other medical imaging modalities. Consequently, models need to be trained for each specific device, which entails substantial costs and time, leading to various practical challenges. To address these challenges, we propose a robust and accurate segmentation network that can operate independently of the ultrasound equipment. We integrated the Deep Frequency Filtering (DFF) module into a U-Net-based model. The proposed model retains the U-Net’s encoder-decoder structure but applies frequency filtering within the latent space of each encoder layer, enabling adaptive selection of frequency components for breast tumor detection. Moreover, batch normalization was replaced with instance normalization to remove stylish features. We evaluated the model using three public datasets acquired from different scanners, achieving superior performance on unseen testing datasets compared to existing models. Notably, when tested on the unseen BUS-BRA dataset, DAUS-Net achieved a Dice score of 0.76, compared to 0.61 by the conventional U-Net. This improvement is attributed to the synergy between the DFF module and instance normalization. Our results demonstrate that the proposed model consistently detects and segments breast tumors, highlighting its potential for generalized clinical segmentation task. The source code for implementing DAUS-Net is publicly available at https://github.com/shlee8638/DAUS-Net.