Abstract: Contrast-enhanced ultrasound (CEUS) is a vital technique for lesion identification, where accurate localization and segmentation of target regions are essential tasks. However, extracting reliable quantitative information from CEUS data is significantly challenged by respiratory motion artifacts that disrupt spatial consistency and the inherent fuzziness of lesion boundaries. Existing methods generally use a single-modal signal, i.e., the fundamental or harmonic signal, in the data, failing to leverage the complementary information from multimodal signals. In this work, we developed a bimodal deep learning network for fusing the bimodal features. The model synergistically fuses fundamental and harmonic signals by integrating residual learning and depth-wise separable convolution (DSCNN). The fundamental signal provides anatomical structural details, while the harmonic signal captures dynamic perfusion profiles. Identification of focal liver lesions (FLLs) was performed with a group of CEUS images. The pixel-level fusion strategy effectively enhanced signal resolution and robustness against motion interference, enabling the proposed model to outperform the established models, such as standard U-Net, fully convolutional network (FCN) and Attention U-Net. This work may provide a powerful tool for rapid and accurate FLL identification, offering a technique for processing multimodal signals.

Key words: Deep learning, Multimodal signal fusion, Depth-wise separable convolution, Focal liver lesion