High-Fidelity Visualization of Large Medical Datasets on Commodity Hardware

Recent advances in CT and MRI static and dynamic scanning techniques have led to great improvements in the resolution and size of volumetric medical datasets, and this trend is still ongoing. However, the explosion of dataset size prevents clinicians from taking advantage of an interactive, high-resolution exploration of volumetric medical data on commodity hardware, due to the memory constraints of modern graphics cards. This paper presents a hybrid CPU-GPU volume ray-casting method and some hybrid-based inspection tools aimed at providing interactive, medical-quality visualization using an ordinary desktop PC. Experimental results show that the hybrid method provides a near-interactive high-fidelity visualization of large medical datasets even if only limited hardware resources are available. 1. Introduction Over the past few years, direct volume rendering (DVR) has stood out as a powerful means to visualize volumetric medical data. The main benefit of volume rendering techniques is that they allow you to capture all the 3D data in a single 2D image, conveying much more information than indirect rendering techniques [1]. Nonetheless, only recently we have been experiencing a broader adoption of DVR techniques in daily clinical practice. For a long time, insufficient image quality [2] and poor interactivity [3] had been the main barriers to the adoption of volume visualization in the clinic. Medical visualization requires interactive response times in image generation, but until a few years ago interactivity was achievable only if high-end, highly expensive graphics workstations were used. The turning point was the dramatic increase in the programmability and computational precision of modern graphics processing units (GPUs), which have made it possible to visualize volumetric medical data with a high degree of accuracy and at an interactive frame rate. Vertex and fragment units of the new generation of GPUs are now user programmable, allowing applications to take advantage of their massive many-core architectures to speed up data parallel tasks. Moreover, whereas few years ago graphics hardware supported only 32-bit color and so provided only 8-bit values to store the voxel intensities, now the GPU’s pixel depth reaches 128 bits per pixel, which means that each component (red, green, blue, and alpha) has a 32-bit floating point precision throughout the graphics pipeline. As a consequence, in recent years much research activity has been concentrated on the design of GPU-accelerated DVR techniques suitable for the visualization in real time of 3D [4–7]