Benign prostatic hyperplasia (BPH) is one of the main diseases affecting the health of middle-aged and elderly men. The accurate measurement and reconstruction of the surgical region (SR) from medical images is a vital and challenging step before the surgery. In this article, an automatic prostate SR reconstruction method based on multilevel learning is proposed. This method divides the reconstruction problem into two levels: prostate segmentation learning and reconstruction parameter learning. It can not only segment the prostate accurately, but also fuse various surgical constraints flexibly and doctors' clinical experience. Compared with traditional methods, the proposed method has better reconstruction accuracy and flexibility. The proposed method was comprehensively validated on multiple datasets. It achieved better accuracy and robustness than current baselines on the magnetic resonance (MR) images of 20 patients from public and clinical datasets. Moreover, the clinical patients' postoperative MR images were collected, and a preoperative-postoperative comparative study was carried out, which further proved the effectiveness of this method from a clinical perspective. Furthermore, this method has the potential to promote the development of BPH robotic surgery navigation and autonomy as well as improve the safety and efficiency of BPH surgery.