Minimal Information Requirement And Optimization Of Artificial Vision For Visual Prosthesis

Retinitis pigmentosa (RP) and age related macular degeneration (AMD) aremajor eye diseases that cause continual loss of vision and eventuallyblindness. The implantation of a visual prosthesis has been proposed as a wayof providing visual information to blind individuals through the electricalstimulation at various points along the visual pathway, that is, the visualcortex, retina or optic nerve. Usable information is provided to patients in theform of phosphenes, a term used to describe the punctate photic sensationsexperienced after electrical stimulation. Due to the limited number ofelectrodes eliciting limited discrete phosphenes, one of the most importantproblems is how to provide the most useful visual information to prosthesiswearers. This study focused on the minimal information requirement andoptimization of artificial vision based on the interdisciplinary study ofcognitive science and information science.First, we evaluated the simulated phosphene characteristics (positioningand size) based on the tactile perception. The results showed that with theincrease of eccentricity, the positioning average error and dispersion wereincreased. The performances under large, stable and high luminanceconditions were better than those under small, flickering and low luminanceconditions, respectively. Repeated training sessions were shown tosignificantly improve the positioning performance. For the size judgment,there was a linear increase in reported size as visual stimulus size increased.Judgment was significantly affected by stimulus luminance and eccentricity, but not by flicker rates. Brighter and more eccentric stimuli were perceived asbeing larger. Therefore, it suggests to prosthetic stimulation parameters shouldbe biased toward resulting estimation accuracy of phosphene characteristics.On the basis of the study of phosphene characteristics, we optimized themodel of simulated prosthetic vision and studied the effect of phosphenecharacteristics (resolution, distortion, dropout and gray level) and imageprocessing methods (merging pixels to lower the resolution and edgeextraction before lowering resolution) on the objects recognition. The resultsshowed that the objects recognition accuracy reached>90%for32×32resolution and was significantly affected by the resolution, distortion anddropout percentage. With the decrease of the resolution or the increase ofdistortion level and dropout percentage, the recognition decreasedconsiderably; when distortion level>0.4or the dropout percentage>40%,the accuracy was less than60%and objects would be unrecognizable. Theaccuracy decreased significantly only at gray level2while the other threegray levels showed no obvious difference. When a high degree of distortion orpixel dropout, the two methods of image processing had significant difference,and image complexity also significantly affected the objects recognition.To explore the minimum requirement of Chinese reading under simulatedprosthetic vision, we studied the Chinese characters recognition andparagraphs reading and evaluated the effect of phosphene characteristics(distortion, dropout percentage, dot size, gray level or luminance) on theperformance of recognition and reading. Meanwhile, we focused on thedifferent display methods of Chinese characters under simulated irregularphosphene maps to improve the performance of characters recognition andparagraph reading. The results showed that the recognition accuracy ofpixelized Chinese characters was significantly affected by distortion, dropoutpercentage and dot-size change, while gray level had less impact on thecharacters recognition.80%recognition accuracy of Chinese characters couldbe maintained if the distortion level≤0.2, dropout percentage≤20%or pixelsize variability≤4. When reading paragraps, the reading accuracy and efficiency increased with an increase of the resolution of single character in aparagraph. When dropout percentage or gray level increased, the readingaccuracy and efficiency both decreased. The luminance just influenced thereading efficiency. Compared with pixelized character recognition, pixelizedChinese paragraph reading achieved higher accuracy. Optimal Chinesereading performance may require prostheses with about1000electrodes orno more than10%dropout. No matter which display method of Chinesecharacters, distortion of phosphene maps had a significant impact on theperformance of recognition and reading. For less distorted phosphene maps,characters with the electrode array corresponding method had a greatperformance of the recognition and reading, while for more distortedphosphene maps, nearest neighbor search method had a better performance.The nearest neighbor search method was significantly affected by the searchrange and the optimal range was different between phosphene maps withdifferent distortion levels.In general, these findings could lay the theoretical foundations forinformation processing and coding of visual prostheses and contribute to thedesign and development of visual prostheses to provide functional vision forthe blind that will be useful during their daily life.