Efficient MIP volume rendering via fast SIMD interpolation and memory access reordering

This study describes an acceleration method that can perform efficient maximum intensity projection (MIP) visualization, which is essential for medical imaging systems. The proposed method is based on shear-warp volume rendering and produces rendering images using trilinear interpolation in real time without a GPU. This study includes two acceleration methods. First, we propose a high-speed interpolation method using AVX2, which is a single instruction, multiple data system of modern CPUs. Trilinear interpolation can be performed rapidly using the AVX2 instructions by taking advantage of the fact that each interpolation weight is the same while using the shear-warp volume rendering. Second, we propose a method for efficiently accessing the memory, focusing on the fact that changing the order of the comparison operations does not affect the image quality in MIP. We propose a new method for changing the repetition and memory access orders so that large volume data can be read sequentially, and image data can be accessed repeatedly. Moreover, we investigate the effectiveness of aligned memory access. The experiment demonstrates significant improvements compared to existing methods. As a result, volume data composed of more than 500 images used in clinical practice can be rendered in real time using trilinear interpolation. In this study, high-quality MIP volume rendering is possible in real-time with only CPU. Since this study does not go through a complicated pre-processing process, it can be easily applied to existing medical imaging systems.