Abstract: Existing interrupted sampling repeater jamming (ISRJ) suppression methods often rely on explicit jamming parameter estimation, insufficiently model the amplitude-phase coupling of complex IQ echoes, and inadequately characterize the long-range structure caused by delayed and repeated forwarding. To address these limitations, a complex-domain end-to-end target echo extraction method is proposed. The jammed complex IQ echo is used as the input, and the ideal target echo is used as the supervision signal, so that ISRJ suppression is formulated as a complex target echo recovery task. The proposed CV-LSTM-UNet adopts a U-Net encoder–decoder architecture to extract multiscale features. Complex convolution is employed to preserve the amplitude-phase correlation between the in-phase and quadrature components. A unitary-constrained channel mixing mechanism is introduced to improve the stability of deep complex feature propagation. In addition, a complex bidirectional temporal modeling module is embedded in the bottleneck layer to characterize the long-range dependence induced by delayed copying, repeated forwarding, and cyclic forwarding. Simulation results show that the proposed method achieves the lowest PSLR and ISLR in three typical ISRJ scenarios, including direct forwarding, repeated forwarding, and cyclic forwarding. In the cyclic forwarding scenario, the PSLR and ISLR reach -35.94 dB and -22.33 dB, respectively. Ablation experiments verify the effectiveness of complex bidirectional temporal modeling and unitary-constrained channel mixing. Robustness and generalization experiments further indicate that the proposed method maintains good adaptability under different SNR/JSR conditions and some unseen jamming parameters. The results demonstrate that the proposed method can effectively suppress ISRJ-induced coherent false peaks while preserving the complex amplitude-phase structure of the target echo.