摘要： Objective: To determine the limits of the optic nerve head (ONH) in color fundus images using Deep learning (DL) for the estimation of its hemoglobin topographic distribution. Also, to evaluate the usefulness of that distribution in glaucoma diagnosis singly or in association with perimetry. Methods: A DL method was trained using 40000 fundus images and applied to 89 normal eyes and 77 confirmed or suspect glaucomas. DL and manual segmentation were compared. The eyes were also examined once with TOP perimetry (Octopus 300) and Spectralis-OCT and twice with Cirrus-OCT and Laguna ONhE, a program which estimates hemoglobin from color photographs, using improved criteria from previous studies. Results: The Sorensen-Dice similarity index between manual and automatic segmentations was 0.993. Intra-class correlation coefficients were similar when comparing the results of the Laguna ONhE indices using the manual and automatic segmentations (confidence intervals: 0.933-0.978). For specificity close to 95%, the GDF index, a factor that measures the distribution of hemoglobin at the nerve, obtained sensitivities between 70.1 and 74.0% (manual vs. automatic segmentations). The retinal nerve fiber layer thickness (RNFLT) of both OCTs provided sensitivities between 67.1 and 68.8% and the BMO-RMW of Spectralis-OCT 69.7%. Associating several normalized indices, e.g. a new visual field harmony index (Threshold Coefficient of Variation, TCV) and GDF, provided 85.7% sensitivity for 97.8% specificity. GDF correlation with Spectralis-OCT BMO-RMW index was similar to that obtained between this index and the RNFLT of the same instrument. For 95% specificity, the diagnostic concordance (kappa value) between both Spectralis-OCT indices was 0.694 and between its BMO-RMW and Laguna ONhE GDF 0.804-0.828. Conclusion: A fully automatic delimitation of the optic nerve head allows the correct, reproducible and efficient use of the Laguna ONhE method, and its effectiveness is greatly increased if associated with a perimetric harmony index.