Abstract: In ultrasound tomography based on full-waveform inversion (FWI), the lack of low-frequency data, due to the conventional medical ultrasound probes having frequencies greater than 1 MHz, makes it difficult for traditional algorithms to reconstruct high-quality images. This paper proposes a deep learning approach for low-frequency extension of ultrasound tomographic scan data, which improves the reconstruction accuracy of FWI by predicting the missing low-frequency signals. An ultrasonic bandwidth expansion network (UBE-Net) is designed and implemented to achieve an end-to-end mapping from the broadband ultrasound signals collected by the detector to the low-frequency extended signals. The network was trained using publicly available simulated datasets and our own phantom datasets, and its performance was tested on simulated, phantom, and in vivo datasets. Experimental results demonstrate that the proposed method effectively enhances the low-frequency components of the detector data, with the signal-to-noise ratio (SNR) of the band-extended signals being higher than 25 dB and 22 dB in simulated and phantom experiments, respectively, and the normalized mean square error being less than 0.07. Using the band-extended ultrasound data for FWI significantly improves the quality of the reconstructed acoustic velocity distribution maps. In simulated and phantom experiments, the peak signal-to-noise ratio of the acoustic velocity distribution maps reached at least 34 dB and 29 dB, respectively, and the structural similarity index was higher than 0.81. Tests on in vivo datasets indicate that the natural image quality evaluator of the reconstructed acoustic velocity distribution maps is below 5.3, and the no-reference structural similarity is higher than 0.79.