Fig. 1: Performance of the quality assessment module.

a In the external test dataset, the assessment module achieved an AUC of 0.933 for distinguishing the location of fundus images. For posterior images, the assessment module achieved AUCs of 0.952–0.995 for detecting various quality defects. For peripheral images, the assessment module achieved AUCs of 0.975–0.989 for detecting various quality defects. b For each of the ten models in the quality assessment module, the original images (left) and corresponding heatmaps (right) are shown. The heatmaps show that the models are focused precisely on the exact regions with quality defects. The correlation matrix shows the Spearman’s rank correlation coefficients between the IQCS ranking, individual experts’ ranking, and consensus ranking in posterior images (c) and peripheral images (d). e, f Distributions of IQCS in infantile fundus images with different quality grades. e For posterior images, the results are shown for 589 excellent quality images, 446 eligible quality images, and 110 ineligible images. f For peripheral images, the results are shown for 573 excellent quality images, 5776 eligible quality images, and 1015 ineligible images. The dashed lines represent the hypothetical thresholds. Post posterior, Peri peripheral, IQCS image quality comprehensive score, Con. consensus, Ex. experts.