Development of an artificial intelligent grading diagnosis model for diabetic fundus lesions based on EasyDL and its verification evaluation

Abstract

Objective To innovatively utilize the open artificial intelligence （AI） platform EasyDL to independently develop an AI auxiliary diagnosis model for diabetic retinopathy （DR）, and evaluate its diagnostic accuracy indicators. Methods 35 126 fundus photos of the diabetes fundus disease data set published by Kaggle were used as the training set, and uploaded to the EasyDL open platform to establish an AI auxiliary diagnosis model. A total of 300 color fundus photographs of bilateral eyes of 150 patients with diabetes mellitus who received clinical DR screening were collected as the test set. The diagnosis of 3 ophthalmologists with deputy director title or above was considered as the gold standard. The diagnostic accuracy for the grading of DR by the AI diagnosis model, junior physicians, intermediate physicians and these combined was evaluated, respectively. Results There were 170 patients with non-DR （NDR） and mild non-proliferative DR （NPDR）, and 130 patients with moderate and severe NPDR and proliferative DR （PDR）. AI diagnostic model had high sensitivity but low specificity. AI diagnostic indexes were close to those of intermediate doctors and better than primary doctors. When AI diagnostic model was combined with physician diagnosis, the accuracy and sensitivity of diagnosis were improved. In the consistency evaluation with the gold standard, the Kappa coefficient of the AI diagnosis model was 1.00, and 0.88 for the intermediate physicians （both P < 0.01）. Conclusions The AI diagnosis model based on the open platform EasyDL is simple and easy to operate, which can contribute to the preliminary screening of DR. It also provides effective scientific research tools for physicians who lack of the knowledge of deep learning algorithms.

Keywords

• |diabetic retinopathy|artificial intelligent diagnosis model|diagnostic accuracy study