Automated Assessment of Choroidal Mass Dimensions Using Static and Dynamic Ultrasonographic Imaging.

PURPOSE: The purpose of this study was to develop and validate an artificial intelligence (AI) model for detecting and measuring choroidal mass dimensions on B-scan ophthalmic ultrasound images.

METHODS

The study included 1822 still images and 130 cine loops of choroidal masses. For external validation, 180 additional still images were included, along with 374 control images to assess specificity. A two-stage U-Net-based architecture was trained to detect masses and measure dimensions. For cine loops, the algorithm automatically selected the frame with the largest mass cross-sectional area.

RESULTS

In the internal subset, detection accuracy was 94.5% with a false-positive rate of 11.7%. For apical height, the mean absolute error (MAE) was 0.42 ± 0.58 mm (R² = 0.87), with 94.2% of cases within 1 mm of expert annotations. For basal diameter, the MAE was 1.02 ± 0.99 mm (R² = 0.74). In the external validation subset, detection accuracy was 83.9%, false-positive rate of 4.2%, and consistent millimeter-level precision for both apical height and basal diameter. Among cine loops, masses were detected in 99.2% of cases, with spatial awareness in 93.1%. Best-frame analysis yielded apical height within 1 mm of the reference in 68.2% of cases (MAE 1.10 ± 1.36 mm) and basal diameter with an MAE of 1.65 ± 1.84 mm.

CONCLUSIONS

Deep learning provides reproducible, millimeter-level measurements of choroidal mass dimensions from still images and cine loops, supporting its potential use for monitoring of choroidal tumors.

TRANSLATIONAL RELEVANCE: AI can generate precise measurement of choroidal tumors, enabling clinically actionable monitoring from ophthalmic ultrasound.