Deep learning-based no-reference quality assessment of anterior segment ultrasound biomicroscopy panoramic images

Ultrasound biomicroscopy (UBM) enables high-resolution imaging of the anterior segment, essential for accurate diagnosis and anatomical assessment. However, the quality evaluation of UBM currently relies on subjective judgment, which is time-consuming and inconsistent. This study proposes a deep learning (DL)-based no-reference method for objective and automated UBM image quality assessment (IQA), facilitating reliable selection of high-quality images for clinical use.

A total of 1,154 clinical panoramic UBM images of the anterior segment were collected from Tianjin Eye Hospital. The YOLOv8s_DW_FOCUS model was employed to accurately extract the region of interest (ROI) and identify five key anatomical landmarks: the central corneal epithelium, central corneal endothelium, posterior lens capsule, left ciliary groove, and right ciliary groove. In collaboration with clinical ophthalmologists, eight key criteria for assessing anterior segment UBM image quality were established, integrating general medical image evaluation parameters and ophthalmic expertise. Based on these criteria, each frame was assigned a quality score. Images scoring 7 or higher were classified as high-quality, whereas those receiving a perfect score of 8 were considered standard. To validate the feasibility of our method, we conducted rigorous evaluations of its accuracy, focus on key regions, generalization capability, inter- and intra-class discrimination, and consistency in assessment results.

The target detection model achieved a mean average precision (mAP) of 0.935, a recall of 0.898, and a precision of 0.925. Additionally, it effectively focused on key regions, as demonstrated by the heatmap analysis. The t-distributed stochastic neighbor embedding (t-SNE) plot further highlighted the model's strong discriminative capability across different classes and its excellent generalization performance. To assess the consistency between our no-reference quality assessment method and expert evaluations, we analyzed 174 standard images that had been subjectively selected by clinical ophthalmologists. Among them, 146 images received a score of 8, whereas 26 images scored 7, indicating a high level of agreement with clinical experts in identifying high-quality images. Moreover, our method applies stricter criteria for defining standard images, enabling a more precise selection of high-quality anterior segment UBM images.

The DL-based no-reference quality assessment method proposed in this study provides an objective evaluation of anterior segment UBM image quality. It effectively identifies high-quality images, significantly improving the efficiency of ophthalmic imaging professionals and demonstrating strong clinical potential for widespread adoption.