Research On Automatic Grading And Focus Identification System Of Myopic Macular Degeneration Based On Fundus Color Photographs

PurposeTo develop deep learning models based on color fundus photographs that can automatically grade myopic maculopathy,diagnose pathologic myopia,identify and segment related lesions,so as to serve the purpose of screening,diagnosing and monitoring disease progression.Methods895 patients with myopia,and 1395 macula-centered 45-degree posterior pole color fundus photographs were included.The definition,grading and lesions annotating criteria of pathologic myopia were based on the grading system proposed by the meta-analysis for pathologic myopia study group in 2015.All photographs were graded and annotated by a team composed of 4 ophthalmologists.Each photograph was annotated by at least 2 ophthalmologists independently.Afterwards the annotation results were discussed and reviewed by a senior ophthalmologist,and were then divided into a high consistency subgroup or a low consistency subgroup according to the consistency between the results.The partition ratio of the training set,validation set and test set in the high consistency subgroup was approximately 60%,20%and 20%.A ResNet-50 network was used to develop the classification model,and a DeepLabv3+network was used to develop the segmentation model for lesions identification.Accuracies and quadratic-weighted ? coefficients were calculated to evaluate the consistency between the model and the reference;an area under the receiver operating characteristic curve(AUC),sensitivity,specificity and Youden index were calculated to evaluate the ability of the model to diagnose pathologic myopia;precisions,recalls and F1 values,etc.,were calculated to evaluate the ability of the model to identify and segment the lesions.Results1.The grading accuracy of the basic classification model on the test set of the high consistency subgroup(1203 photographs)was 0.9076,and the quadratic-weighted ? coefficient was 0.9324;the grading accuracy of the classification-and-segmentation-based co-decision model was 0.9370,and the quadratic-weighted ? coefficient was 0.9651;the co-decision model achieved an AUC of 0.9980,with a sensitivity of 0.9667 and a specificity of 0.9915 in diagnosing pathologic myopia.In the low consistency subgroup(192 photographs),the grading accuracy of the co-decision model was 0.5677,and the quadratic-weighted ? coefficient was 0.7640;if the two grades of a photograph,between which the ophthalmologists had a disagreement,were both regarded as the ground truth,then the grading accuracy would be 1.0000.2.The photograph-level F1 values of the segmentation model identifying optic disc,peripapillary atrophy,diffuse atrophy,patchy atrophy and macular atrophy were all>0.95,and the pixel-level F1 values of the model segmenting optic disc and peripapillary atrophy were both>0.9,and the pixel-level F1 values of the model segmenting diffuse atrophy,patchy atrophy and macular atrophy were all>0.8;the photograph-level recall of the segmentation model identifying lacquer cracks was 0.9230;the average optic disc ovality index of each grade calculated according to the segmentation model changed in parallel with that calculated according to manual annotation.ConclusionThe deep learning models based on color fundus photographs could accurately and automatically grade myopic maculopathy,diagnose pathologic myopia,and identify and monitor the progression of peripapillary atrophy,various atrophic lesions,lacquer cracks,and tilted optic disc.The models can provide effective help releasing the burden of diagnosis,screening,and follow-up in the prevention and control of myopia.