Introduction: Myocardial perfusion imaging (MPI) is a reference technique for coronary artery disease (CAD) diagnosis. This paper focuses onto 3D PET image processing prior to quantification. For visualization and quantitative image analysis, the transaxial reconstructed tomographic images have to be properly rotated to the standard horizontal-long axis, vertical-long axis and small-axis views. This specific 3D rotation step induces an interpolation and the resulting degradation must be carefully quantified so that it can be minimized. Material and methods: Continuous rotation on a discrete grid introduces bias in rotated data due to the lack of perfect match from the initial grid to the rotated one. It follows that a discrete 3D angle description (defined through integer displacements along the x, y and z axes) can always be sufficient to assess the rotation. A method for the error quantization of a rotation method is successive rotations followed by inverse rotation of the opposite angle. The resulting image is different from the original one and the distance between them is used as an indicator of the algorithm performance. We have investigated and compared two classes of rotation algorithms: (classical) continuous rotation with bilinear interpolation versus discrete rotation algorithm based on Finite Radon Transform (FRT). The FRT is a discrete and exact form of the Radon transform using discrete angles. The underlying idea is to rotate the data not in the image space, but in the discrete projection space which offers an intrinsic rotational structure. Since the FRT is exact, the inverse transform is also exact and no information is lost between the projection and the backprojection of rotated data. FRT based rotations allows for permuting 2D projection data to implement a 3D rotation, which requires significantly less computation. This provides us fast and accurate reorientation algorithm. Results: Tests were conducted both onto a 3D PET FDG cardiac volume and on synthetic data. The L2 distance between original and rotated data was computed inside a volume of interest. Any 3D continuous rotation can be decomposed into two discrete 2D angles, with arbitrarily close precision, through a ratio of integers chosen from the Farey-Harros series. This technique is compared to classical interpolation with or without data oversampling in the final paper. Conclusions: We propose a discrete, 2D projection-based operator to achieve precise, high quality rotation of 3D tomographic data. Our algorithm compares favorably to the continuous rotation using interpolation of adjacent voxels.