External Validation of Deep Learning OCTA Quality Control for Foveal Avascular Zone Analysis

Background and Aim : Image quality in optical coherence tomography angiography (OCTA) can limit quantitative analysis of the foveal avascular zone (FAZ). We developed an automated image-level quality-control (QC) pipeline and assessed its performance on an external test set from Erasmus MC. Methods : An ImageNet-pretrained DenseNet-161 was fine-tuned on labelled OCTA images from Jules-Gonin Eye Hospital (HOJG) to classify images as high or low quality. Training used image augmentation, class-weighted cross-entropy and validation-based checkpoint selection. External validation used 22 Erasmus MC images, with no fine-tuning on Erasmus MC data. Performance was measured using the area under the receiver operating characteristic curve (AUROC), macro-averaged F1-score (macro-F1), precision and recall, accuracy, and specificity. Results : On the external Erasmus MC test set, AUROC was 0.7946, macro-F1 was 0.6364 and accuracy was 63.6%. Specificity, defined as correct rejection of low-quality images, was 87.5%. The pipeline exported an image-level quality label and probability for batch image selection. Conclusion : OCTA_QC discriminated between high- and low-quality images in a small external test set without site-specific fine-tuning. The pipeline offers an automated image-selection step before FAZ analysis. Larger evaluations and direct measurement of downstream benefit are needed before broad autonomous use.