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

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Biomedical Engineering

Background:

• Accurate medical image segmentation is vital for diagnosis, but deep learning models require extensive labeled data, which is costly and time-consuming to acquire.
• Ultrasound imaging exhibits significant equipment-dependent feature variations, necessitating device-specific model training and posing practical challenges.

Purpose of the Study:

• To develop a robust and accurate segmentation network for ultrasound breast tumor detection that is independent of specific ultrasound equipment.
• To enhance the generalizability of deep learning models in medical imaging segmentation.

Main Methods:

• Integration of a Deep Frequency Filtering (DFF) module into a U-Net architecture.
• Application of frequency filtering in the latent space of encoder layers for adaptive component selection.
• Replacement of batch normalization with instance normalization to eliminate style-specific features.

Main Results:

• The proposed DAUS-Net demonstrated superior performance on unseen datasets compared to conventional U-Net.
• Achieved a Dice score of 0.76 on the BUS-BRA dataset, significantly outperforming the U-Net's 0.61.
• Consistent detection and segmentation of breast tumors across different scanners, attributed to DFF and instance normalization.

Conclusions:

• DAUS-Net offers a generalized solution for ultrasound breast tumor segmentation, overcoming equipment variability.
• The model's performance highlights the potential for improved clinical segmentation tasks with reduced data acquisition costs.
• Publicly available source code facilitates further research and development in medical imaging AI.