Network architecture optimization for ophthalmic ultrasound image detection based on modular ablation of...

To address the absence of a systematic evaluation method for network architecture selection in ophthalmic ultrasound image detection tasks, this study proposes a modular ablation analysis framework based on orthogonal experimental design.

A clinical data set comprising 1121 ocular ultrasound images was established. YOLOv10-v12 were decoupled into backbone, neck, and head modules. A three-stage evaluation was conducted: (1) single-module benchmarking, performed via controlled variable experiments; (2) orthogonal combination experiments using an L9(3⁴) array, analyzed through range analysis and interaction heatmaps; and (3) optimal architecture selection, implemented via Pareto front analysis. The best model was applied to ocular tissue localization, and a segmented sound velocity matching algorithm was used to automatically measure biometric parameters, including anterior chamber depth, lens thickness, and axial length.

The backbone improved both accuracy and efficiency, while the neck and head exhibited a speed-accuracy trade-off. The neck most significantly influenced detection accuracy, and the head dominated computational efficiency. The optimal combination (Bv11-Nv11-Hv10) achieved 64.0% mAP at 26 FPS, while the mobile-optimized variant (Bv10-Nv10-Hv11) attained 63.5% mAP with only 8.6 MB parameters. Automatic and manual measurements showed strong agreement (mean absolute error ≤ 0.133 mm, ICC ≥ 0.839).

This study validates the feasibility of cross-version module combination. The proposed framework offers a systematic, quantitative decision-making basis for network design in ophthalmic ultrasound, balancing accuracy, speed, and deployment feasibility. Clinical results confirm high consistency between automatic and manual measurements, supporting automated and precise ocular biometry.