Pseudoisochromatic Plates Synthesis Based On Generative Adversarial Network

Pseudoisochromatic plates are among the most popular tests for color visual deficiency(CVD),which is used to screen and evaluate individual color vision for its convenience compared to other color vision diagnostic methods.At present,most of the pseudoisochromatic plates used in clinic are pre-drawn fixed books,which is the main tool for the qualitative assessment of color vision.However,with the continuous improvement of medical and health industry,and the continuous development of clinical needs driven by precision medicine,a more accurate and detailed assessment method of human color vision should be further investigated.Aiming at this point,this thesis proposes two new methods for synthesizing pseudoisochromatic plates by using traditional image processing and deep learning way respectively.This thesis mainly completed the following works.In this paper,firstly by the aid of color vision simulation method,a synthesis method of pseudoisochromatic plates based on pseudoisochromatic principle is proposed.Assuming that Brettel et al.'s simulation model is qualitatively correct,which can be used to compute the corresponding pair of arbitrary given color with respect to three type of dichromacy,confusion colors among perceptual color space are calculated,obtained and persisted by means of grid search and color difference threshold selection.Then the elements in the obtained color set are used as the central point to filling the pattern blocks in image samples with their nearby colors within a sphere.Finally,k-d tree is introduced to replace Monte Carlo method to accelerate the drawing procedure of non-overlapping circles.On the basis of synthetic method of pseudoisochromatic principle,this paper further studies the synthetic method based on conditional generative adversarial networks.The generator uses an end-to-end structure resembles the U-net,while the discriminator builds with only full convolutional neural network follows the manner of PatchGAN.Then,the input 2D noise,content mask,color vector and targeted CVD type information are fused through two-stage operation to adapt the generator.During the training procedure,pairwise dataset is constructed using the synthetic method mentioned above,and the simulation loss function and specificity loss function are introduced to improve the fitness of the network.In order to verify the effectiveness and feasibility of two image synthesis methods,evaluation metrics were devised utilizing the segmentation mask of salient foreground before and after simulation.The samples produced by each method are analyzed with the aspect of compliance,specificity and robustness,as well as time complexity.From the experimental results,it can be concluded that the pseudoisochromatic plates synthesized by both methods are of high compliance and specificity,which the GAN-based synthesis shows slightly higher robustness.The score of compliance and specificity after removing the abnormal samples are about 0.901 and 0.896 for the first method,and 0.923 and 0.917 for the GAN based counterpart,which indicates that the latter are at least 2%higher that the former.The overall work has proved the effectiveness and potential usage of our methods,which may promote the computer-aided diagnosis for human color vision deficiency.