The registration of 2D-3D medical images is crucial in solving radiotherapy positioning verification. A 2D-3D medical image registration approach based on training-inference decoupling architecture is proposed to address the issues of low accuracy and time-consuming processes. The multibranch structure and multiscale convolution were employed in the training phase to enhance feature diversity and improve registration accuracy. During the inference phase, the multibranch structure was reparameterized into a single-channel structure to speed up the registration speed. Additionally, an adaptive activation function, Meta-ACON, was used to increase the network’s nonlinear expression. Two datasets of the chest and pelvis were used for training and testing. The experimental results show that the mean translation error of the proposed method is approximately 0.08 mm, the mean angular error is approximately 0.05°, and the registration time reaches 26 ms. The proposed method significantly improves the accuracy of medical image registration in positioning verification while meeting the real-time requirements of clinical applications.