Automatic Diagnosis of Plus Disease in Retinopathy of Prematurity Using Deep Learning

OBJECTIVES: Plus disease, which can be diagnosed during clinical ophthalmoscopic examinations, is the most important feature in determining retinopathy of prematurity (ROP) requiring treatment. In the current study, we aimed to automatic diagnose and predict plus disease based on deep learning (DL) from retinal images of infants with ROP.
METHODS: 600 retinal images from infants screened for ROP were evaluated. Each image was classified as normal, pre-plus and plus disease. After image pre-processing, the images were distributed into groups equal to the number of normal eye images and were used for training DL algorithms such as EfficientNetB7, InceptionResNetV2 and VGG16. The algorithms were trained with 10-fold cross-validation, and results are reported as sensitivity, specificity, accuracy, receiver operating characteristic (ROC) curve, and area under the curve (AUC).
RESULTS: Of the 600 retinal images included, 258 images obtained after pre-processing were graded as 86 (33.3%) normal, 86 (33.3%) as pre-plus disease, and 86 (33.3%) as plus disease. For plus versus no-plus disease diagnosis, among the algorithms, InceptionResNetV2 achieved 0.90 sensitivity, 0.92 specificity, and 0.91 accuracy. Area under the ROC curve was 0.91. For detection of pre-plus disease or worse versus normal, the algorithm achieved 0.81 sensitivity, 0.97 specificity, and 0.88 accuracy.
DISCUSSION AND CONCLUSION: Our results showed that DL algorithms can automatically diagnose plus disease in ROP with high sensitivity, high specificity, and high accuracy.

Keywords: Deep learning, fundus images, plus disease, retinopathy of prematurity