THREE-DIMENSIONAL (3D) CONVOLUTION WITH 3D BATCH NORMALIZATION

摘要

The technology disclosed uses a 3D deep convolutional neural network architecture (DCNNA) equipped with so-called subnetwork modules which perform dimensionality reduction operations on 3D radiological volume before the 3D radiological volume is subjected to computationally expensive operations. Also, the subnetworks convolve 3D data at multiple scales by subjecting the 3D data to parallel processing by different 3D convolutional layer paths. Such multi-scale operations are computationally cheaper than the traditional CNNs that perform serial convolutions. In addition, performance of the subnetworks is further improved through 3D batch normalization (BN) that normalizes the 3D input fed to the subnetworks, which in turn increases learning rates of the 3D DCNNA. After several layers of 3D convolution and 3D sub-sampling with 3D across a series of subnetwork modules, a feature map with reduced vertical dimensionality is generated from the 3D radiological volume and fed into one or more fully connected layers.

主权项

1. A method of convoluting three-dimensional (3D) data in a deep neural network, the method including:
receiving three-dimensional (3D) data characterizing an input radiological volume; processing the 3D data characterizing the input radiological volume using a deep neural network to generate alternative feature volume representations of the input radiological volume, wherein the deep neural network comprises a plurality of subnetworks, wherein the subnetworks are arranged in a sequence from lowest to highest, and wherein processing the data characterizing the input radiological volume using the deep neural network comprises processing the data through each of the subnetworks in the sequence; wherein three or more of the subnetworks are module subnetworks and each of the module subnetworks is configured to:
receive a preceding output representation generated by a preceding subnetwork in the sequence;process the preceding output representation through at least three parallel 3D convolution layer paths of varying convolution volume;further process the preceding output representation through a parallel 3D max pooling path; andconcatenate output of the 3D convolution layer paths and the 3D max pooling path to generate an output representation from each of the module subnetworks; and following processing the data through the three or more module subnetworks, processing the output of the highest of the module subnetworks through a vertical max-pooling layer to generate an output of reduced vertical dimensionality from the input radiological volume.