Fig. 2: Performance of the quality enhancement module.

a Comparison of IQCS in infantile fundus images before and after quality enhancement. Each line represents the change in IQCS before and after enhancement of one fundus image. b, c Typical examples of raw fundus images and the corresponding enhanced fundus images for ROP. Compared to the raw image, the corresponding enhanced image demonstrates better visibility of retinal lesions and vessels. b White arrows indicate peripheral retinal ridge-like elevation lesions, which suggest a stage 2 ROP lesion. c Red dashed circle indicates that the retinal vessels were tortuous and dilated, which suggests plus-disease of ROP. d Clinicians were assigned to diagnose ROP according to 100 raw images and corresponding enhanced images. The accuracy, sensitivity and specificity were compared. The p value was calculated using the independent two-sample t-test. The error bars represent the standard deviation of the clinicians in the corresponding groups. The performance of ROP diagnosis models developed and evaluated by raw images and corresponding enhanced images using InceptionV3 (e) and DenseNet (h) were compared. t-Distributed stochastic neighbor embedding visualization of features extracted from an intermediate layer of trained models for ROP diagnosis using InceptionV3 (f, g) and DenseNet (i, j) architectures. This visualization demonstrates the capacity of different models to distinguish ROP images and normal images. Orange points represent normal images, and blue points represent ROP images. There was greater intergroup distance and lower intragroup distance in models with the quality enhancement module, which indicated greater performance for ROP diagnosis. ROP retinopathy of prematurity. **p < 0.01; ***p < 0.001.