Multi-objective Ranking to Optimize CNN’s Encoding Features: Application to the Optimization of Tracer Dose for Scintigraphic Imagery
Abstract
The pooling layer is at the core of every convolutional neural network (CNN), contributing to data invariance, variation, and perturbation. It describes which part of the input image a neuron in the output layer can see. CNNs with max pooling can handle simple transformations like flips or rotations without too much trouble. The problem comes with complicated modifications. The rank order importance is used here as an alternative to max-pooling. The rank texture descriptor is non-parametric, independent of geometric layout or size of image regions, and can better tolerate rotations. These description functions produce images that emphasize low/high frequencies, contours, etc. We propose a multi-objective ranking algorithm derived from Vargas et al. [10] to optimize CNN’s encoding features. It is applied for the first time to estimate trace dose in radiology with scintigraphic imagery.