As a key technology to evaluate cement bonds in the cased hole, an advanced ultrasonic logging tool combines pulse-echo and pitch-catch measurements in which the latter one provides reflections from the cement-formation interface (called third-interface-echo [TIE]) to evaluate the bond condition and determine casing eccentering as well as cement velocity. However, the TIE would be weak and not easy to pick due to the eccentered tool and casing and it would overlap with the strong multiple reflections between the casing inner surface and the transducer-housing tool. We have developed a deep learning workflow to extract weak TIE from noisy data and to preserve its amplitude at the same time. First, we use synthetic waveforms from thousands of finite-difference simulations as initial training data sets to train a deep learning network, which is modified from a network in speech separation. Then, the trained model is used to predict the field data through an active-learning strategy. The improved network is further used to extract the weak TIEs, which are not easy to pick in the initial deep learning model. Finally, the TIE waves image is converted to a pseudovelocity image to obtain the minimum traveltime path by solving the eikonal equation. The shortest traveltime path is used as the TIE arrival time. In addition, a 3D visualization is used to display the borehole shape from the picked arrival time. The applications in synthetic data and data set from a calibration well illustrate a good performance of our workflow in which the weakest TIE extracted from the network can reach 50 dB compared to the maximum amplitude in the full waveform. The picked arrival times can be used to reconstruct a borehole shape.