A Supervised Generative Model for Efficient Rendering of Medical Volume Data

Abstract

Complex 3D and multidimensional medical data require significant computational resources to render high-quality, high-resolution images which provide useful information from the originating data set. As an alternative to traditional on-the-fly ray-casting-based generation of such images from volume data, we propose a generative model based on a deep neural network which is continually trainable from data dynamically generated by a GPU-based renderer. When properly trained, the model is capable of independently generating images similar to ones produced by a dedicated renderer, using control parameters commonly encountered in volume rendering engines. The network takes over the task of generating images from such parameters, thereby alleviating the need for high-capability computational resources while at the same time providing images without requiring access to the original data sets. Our model allows the user to generate high-resolution images on low-spec hardware without the need for a GPU-based renderer, and without access to sensitive or protected patient data. Also, the model is exploitable in a manner which allows the fully-interactive exploration of complex volume data sets and the efficient generation of representations of the data using limited computational resources.