Macular Vessel Density and Ganglion Cell/Inner Plexiform Layer Thickness and Their Combinational Index Using Artificial Intelligence.

Abstract

To evaluate the relationship between macular vessel density and ganglion cell to inner plexiform layer thickness (GCIPLT) and to compare their diagnostic performance. We attempted to develop a new combined parameter using an artificial neural network. A total of 173 subjects: 100 for the test and 73 for neural net training. The test group consisted of 32 healthy, 33 early, and 35 advanced glaucoma subjects. Macular GCIPLT and vessel density were measured using Spectralis optical coherence tomography and Topcon swept-source optical coherence tomography, respectively. Various regression models were used to investigate the relationships between macular vessel density and GCIPLT. A multilayer neural network with one hidden layer was used to determine a single combined parameter. To compare diagnostic performance, we used the area under the receiver operating characteristic curve (AUROC). Correlation analyses in all subjects showed a significant correlation between macular vessel density and GCIPLT in all sectors (r=0.27 to 0.56; all Ps≤0.006). The fitness of linear, quadratic, and exponential regression models showed clinically negligible differences (Akaike's information criterion=714.6, 713.8, and 713.3, respectively) and were almost linear. In differentiating normal and early glaucoma, the diagnostic power of macular GCIPLT (AUROC=0.67 to 0.81) was much better than that of macular vessel density (AUROC=0.50 to 0.60). However, when vessel density information was incorporated into GCIPLT using the neural network, the combined parameter (AUROC=0.87) showed significantly enhanced diagnostic performance than all sectors of macular vessel density and GCIPLT (all Ps≤0.043). Macular vessel density was significantly decreased in glaucoma patients and showed an almost linear correlation with macular GCIPLT. The diagnostic performance of macular vessel density was much lower than that of macular GCIPLT. However, when incorporated into macular GCIPLT using an artificial neural network, the combined parameter showed better performance than macular GCIPLT alone.