Quantitative susceptibility mapping (QSM) is an MRI phase-based post-processing technique to extract the distribution of tissue susceptibilities, demonstrating significant potential in studying neurological diseases. However, the ill-conditioned nature of dipole inversion makes QSM reconstruction from the tissue field prone to noise and artifacts. In this work, we propose a novel deep learning-based IR2QSM method for QSM reconstruction. It is designed by iterating four times of a reverse concatenations and middle recurrent modules enhanced U-net, which could dramatically improve the efficiency of latent feature utilization. Simulated and in vivo experiments were conducted to compare IR2QSM with several traditional algorithms (MEDI and iLSQR) and state-of-the-art deep learning methods (U-net, xQSM, and LPCNN). The results indicated that IR2QSM was able to obtain QSM images with significantly increased accuracy and mitigated artifacts over other methods. Particularly, IR2QSM demonstrated on average the best NRMSE (27.59%) in simulated experiments, which is 15.48%, 7.86%, 17.24%, 9.26%, and 29.13% lower than iLSQR, MEDI, U-net, xQSM, LPCNN, respectively, and led to improved QSM results with fewer artifacts for the in vivo data.