Purpose To construct a deep learning (DL) algorithm for automatic prediction of sulcus-to‐sulcus diameter (STS) and distance between STS plane and anterior crystalline lens surface (STSL) from ultrasound biomicroscopy (UBM) images based on YOLOv8 and verify its accuracy and reliability. Methods This study used data from 100 eyes of 100 myopic patients treated with ICL from March 2023 to August 2024. UBM was used for the measurements of the STS and STSL (4 images for each eye). The data set (400 images) was randomly split at the patient level into a train, validation and test sets at the ratio of 8:1:1. The ciliary sulci on both sides and the anterior capsule of the lens in the UBM images were located with the YOLOv8 algorithm, and then the distances were calculated and compared with the manual labeled values and compared against an external expert with ANOVA, the YOLOv8 algorithm was tested in 26 eyes (104 images ) independent UBM data sets. Bland–Altman tests and intergroup correlation coefficients (ICCs) were used to assess the agreement between the labeled and YOLOv8 predicted values. Results The deep learning-predicted STS and STSL demonstrated a high level of accuracy and reduced contouring time (by savings of 98.80% of work time) when compared with manual labeling contours in the testing set and showed a good accuracy when compared with external ophthalmologist manual labeling contours and in the external evaluation. The prediction error of the STS being 3.27 ± 2.01% and STSL being 67.95 ± 140.09% for the YOLOv8 algorithm at testing set, 4.10 ± 3.00 (%), and 49.66 ± 42.73 (%) in the external test set. The ICC was 0.312 between the predicted and labeled STS (P = 0.01) and 0.086 between the predicted and labeled STSL (P > 0.05). Conclusions The deep learning-predicted STS and STSL demonstrated high accuracy and reduced measurement time, which could have a positive impact on the clinical setting. Key messages <Emphasis Type="BoldItalic">What is known</Emphasis>:<UnorderedList Mark="Bullet"> <ItemContent> ICL implantation remains challenging because of difficulties in determining the appropriate lens size. </ItemContent> <ItemContent> There is a wide variation in the values of ciliary sulcus-to-sulcus (STS) diameter measurements. </ItemContent> </UnorderedList> <Emphasis Type="BoldItalic">What is new</Emphasis>:<UnorderedList Mark="Bullet"> <ItemContent> This is the first study to automatically measure the STS-related distance based on YOLOv8 and assess the accuracy compared to the conventional manual labeling. </ItemContent> <ItemContent> The YOLOv8 algorithm proposed advantages in high accuracy, automatic prediction of posterior chamber STS-related parameters from ultrasound biomicroscope images. </ItemContent> </UnorderedList>

A pilot study of deep learning for automatic contouring of sulcus-to-sulcus diameter in ultrasound biomicroscopy

期刊论文
发表日期：

2025年3月22日

查看全文

Xiaohong Zheng,
Xiaokang Li,
Ke Hu,
Jingji Long,
Xingtao Zhou,
Yuanyuan Wang,
Yi Guo,
Ke Zheng

摘要

Purpose

To construct a deep learning (DL) algorithm for automatic prediction of sulcus-to‐sulcus diameter (STS) and distance between STS plane and anterior crystalline lens surface (STSL) from ultrasound biomicroscopy (UBM) images based on YOLOv8 and verify its accuracy and reliability.

Methods

This study used data from 100 eyes of 100 myopic patients treated with ICL from March 2023 to August 2024. UBM was used for the measurements of the STS and STSL (4 images for each eye). The data set (400 images) was randomly split at the patient level into a train, validation and test sets at the ratio of 8:1:1. The ciliary sulci on both sides and the anterior capsule of the lens in the UBM images were located with the YOLOv8 algorithm, and then the distances were calculated and compared with the manual labeled values and compared against an external expert with ANOVA, the YOLOv8 algorithm was tested in 26 eyes (104 images ) independent UBM data sets. Bland–Altman tests and intergroup correlation coefficients (ICCs) were used to assess the agreement between the labeled and YOLOv8 predicted values.

Results

The deep learning-predicted STS and STSL demonstrated a high level of accuracy and reduced contouring time (by savings of 98.80% of work time) when compared with manual labeling contours in the testing set and showed a good accuracy when compared with external ophthalmologist manual labeling contours and in the external evaluation. The prediction error of the STS being 3.27 ± 2.01% and STSL being 67.95 ± 140.09% for the YOLOv8 algorithm at testing set, 4.10 ± 3.00 (%), and 49.66 ± 42.73 (%) in the external test set. The ICC was 0.312 between the predicted and labeled STS (P = 0.01) and 0.086 between the predicted and labeled STSL (P > 0.05).

Conclusions

The deep learning-predicted STS and STSL demonstrated high accuracy and reduced measurement time, which could have a positive impact on the clinical setting.

Key messages

What is known:

ICL implantation remains challenging because of difficulties in determining the appropriate lens size.

There is a wide variation in the values of ciliary sulcus-to-sulcus (STS) diameter measurements.

What is new:

This is the first study to automatically measure the STS-related distance based on YOLOv8 and assess the accuracy compared to the conventional manual labeling.

The YOLOv8 algorithm proposed advantages in high accuracy, automatic prediction of posterior chamber STS-related parameters from ultrasound biomicroscope images.