Deep Learning Segmentation, Visualization, and Automated 3D Assessment of Ciliary Body in 3D Ultrasound Biomicroscopy...

Current clinical practices lack efficient, automated methods for quantifying complex ocular structures in three-dimensional ultrasound data. Manual annotation remains slow and prone to human variability during routine diagnostic procedures. That uncertainty drove the need for computational tools capable of rapid, reliable image analysis. Prior research has shown that ultrasound biomicroscopy provides high-resolution views of the anterior eye segment. However, existing software often fails to distinguish between closely related tissues like the muscle and processes. No prior work had resolved the challenge of segmenting these specific components automatically across large image datasets. This gap motivated the development of advanced algorithms to assist clinicians in ophthalmic assessments. The present study addresses these limitations by applying sophisticated machine learning architectures to standardized ultrasound volumes.

The primary aim of this research was to develop a fully automated approach for segmenting and assessing the ciliary body in 3D ultrasound biomicroscopy images. Clinicians currently struggle with the time-consuming nature of manual ocular image analysis. This project seeks to overcome those bottlenecks by leveraging advanced computational models. The researchers intended to create a system capable of distinguishing between the ciliary muscle and processes. They also aimed to validate the accuracy of these automated measurements against manual expert annotations. Furthermore, the study investigated whether the model could detect structural changes following surgical interventions. By providing a faster, more objective assessment, the team hopes to improve ophthalmic treatment planning. This work addresses the critical need for efficient diagnostic tools in modern eye care.

The investigators designed a fully automated pipeline to process three-dimensional ultrasound volumes of human cadaver eyes. They performed multiplanar reformatting to align every volume precisely with the optic axis. Experts manually annotated the muscle and processes to create a reliable training set. The team trained Deeplab-v3+ architectures using various loss functions to evaluate segmentation performance. They tested these models on thousands of radial and en face images to ensure robust results. The researchers also conducted transscleral cyclophotocoagulation to induce structural shrinkage for validation purposes. They compared automated measurements against manual assessments to verify the accuracy of the software. This review approach integrates computational modeling with traditional anatomical measurement techniques.

Radial images processed with Dice loss achieved the highest mean F1-score of 0.89 for overall ciliary body identification. For three-class segmentation, this method reached mean F1-scores of 0.75 for processes and 0.82 for muscle. The algorithm reduced expert editing time by at least seven times compared to manual methods. Average ciliary body volume measured 99 ± 18 mm3, while the surface area reached 709 ± 80 mm2. Approximately 10.9% of muscle tissue was misclassified as process tissue due to anatomical border ambiguity. Following cyclophotocoagulation, both automated and manual techniques confirmed significant reductions in volume and surface area. These findings demonstrate that the model effectively tracks morphological changes in the eye. The data support the feasibility of using this tool for objective ophthalmic assessments.

The authors demonstrate that automated segmentation significantly reduces the time required for expert image editing compared to manual tracing. Their findings indicate that radial image inputs combined with specific loss functions yield superior performance for identifying ocular tissues. The researchers report that the algorithm successfully detects structural changes following therapeutic interventions such as cyclophotocoagulation. Data suggest that automated volume and surface area measurements align closely with manual assessments performed by specialists. The study highlights that misclassification between muscle and process regions persists due to inherent anatomical ambiguity. These results imply that the proposed framework offers a viable path toward standardized clinical monitoring. The team concludes that their approach holds promise for enhancing treatment planning in ophthalmology. This work provides a foundation for future automated diagnostic tools in eye care.