Dual-head PET system suffers from poor resolution in the dimension perpendicular to the detector heads and vastly increased computations required by the accurate estimation of the system response matrix (SRM). The problem of poor x-axis resolution is caused by missing data owing to the intrinsic geometry flaws of the dual-head PET, which leads to high SRM similarity between the adjacent voxels along the x-axis. To compensate for this problem, our dual-head PET system prototype incorporates a [Formula: see text]-rotating mechanism. In addition, the cubic field of view (FOV) and cubic voxels are adopted to achieve a better geometrical symmetry, which avoids the re-estimation of SRM after rotation. The spatial resolution, sensitivity, noise equivalent count rate, scatter fraction and pileup fraction under different head separation distances were evaluated by a series of simulations, and the results were compared with those of the non-rotation system. The SRM was estimated using the Monte Carlo method and the three-dimensional images were reconstructed using the ordered subsets expectation maximization algorithm. Experimental results reveal a significantly improved x-axis resolution via the proposed rotation mechanism. Our system can achieve x-axis spatial resolution of 0.79 mm, 0.74 mm and 0.72 mm at the center of FOV under the head separation distance of 5 cm, 10 cm and 15 cm, respectively.