Aiming at the difficulty in extracting travel time caused by the uncertainty of the pulse width and signal-to-noise ratio of proton thermoacoustic signals in clinic, a travel time extraction algorithm based on dense network is proposed. The algorithm uses dense blocks instead of traditional convolutional blocks, combines features with different receptive fields, and introduces deep supervision and network pruning mechanisms. It uses labeled proton beam thermoacoustic signal data for learning to extract the required time information. The experimental results show that, compared with other algorithms, the proposed algorithm has higher accuracy and robustness for the extraction of proton thermoacoustic signal travel time, and shows the feasibility of real-time extraction.