Kerr J P, Bartlett E B
Biomedical Engineering Program, Iowa State University, Ames, USA.
Comput Biol Med. 1995 Jul;25(4):393-403. doi: 10.1016/0010-4825(95)00017-x.
While finding many applications in science, engineering, and medicine, artificial neural networks (ANNs) have typically been limited to small architectures. In this paper, we demonstrate how very large architecture neural networks can be trained for medical image processing utilizing a massively parallel, single-instruction multiple data (SIMD) computer. The two- to three-orders of magnitude improvement in processing time attainable using a parallel computer makes it practical to train very large architecture ANNs. As an example we have trained several ANNs to demonstrate the tomographic reconstruction of 64 x 64 single photon emission computed tomography (SPECT) images from 64 planar views of the images. The potential for these large architecture ANNs lies in the fact that once the neural network is properly trained on the parallel computer the corresponding interconnection weight file can be loaded on a serial computer. Subsequently, relatively fast processing of all novel images can be performed on a PC or workstation.
虽然人工神经网络（ANNs）在科学、工程和医学领域有许多应用，但通常局限于小型架构。在本文中，我们展示了如何利用大规模并行单指令多数据（SIMD）计算机训练超大型架构神经网络用于医学图像处理。使用并行计算机可实现两到三个数量级的处理时间改进，这使得训练超大型架构人工神经网络变得切实可行。作为示例，我们训练了多个神经网络来演示从64幅平面图像的64个视图进行64×64单光子发射计算机断层扫描（SPECT）图像的断层重建。这些大型架构人工神经网络的潜力在于，一旦神经网络在并行计算机上得到正确训练，相应的互连权重文件就可以加载到串行计算机上。随后，可以在个人计算机或工作站上对所有新图像进行相对快速的处理。
