Convolutional neural network classifiers can differentiate eyes with primary angle closure disease (PACD) from control eyes on anterior segment swept-source optical coherence tomography (AS-OCT), according to a multi-center cross-sectional study published in Ophthalmology Glaucoma. However, these classifiers are only moderately successful when trying to separate different stages of PACD, according to the report.
Researchers evaluated the performance and generalizability of a convolutional neural network classifier model for identification and stratification of PACD on AS-OCT. They recruited patients from 3 different eye centers across China and Singapore between October 2018 and June 2019.
The research team divided 841 eyes from Glaucoma Service at Zhongshan and Beijing Tongren ophthalmic centers into 170 control eyes (mean age, 58.77 years; 54% women), 488 primary angle closure suspect (PACS) eyes (mean age, 63.04 years; 84% women), and 183 eyes with primary angle closure (PAC)/primary angle closure glaucoma (PACG) (mean age, 59.47 years; 72% women).
They recruited an additional 300 eyes from Singapore National Eye Center as testing dataset, divided into 100 control eyes (mean age, 59.2 years; 48% women), 100 PACS eyes (mean age, 71.6 years; 80% women), and 100 eyes with PAC/PACG (mean age, 62.5 years, 50% women).
All participants underwent standardized ophthalmic examination including gonioscopy and AS-OCT imaging. The researchers used deep learning model Inception-v3 to train 3 different convolutional neural network classifiers: classifier 1 aimed to distinguish control vs PACS vs PAC/PACG; classifier 2 aimed to distinguish control vs PACD; classifier 3 aimed to distinguish PACS vs PAC/PACG.
Each convolutional neural network classifier underwent 5-fold cross validation and were assessed on independent test sets from the same region, China. To evaluate the generalizability of these models across different groups of patients, trained classifiers were further tested utilizing data from a different country, Singapore.
Classifier 1 showed an area under the receiver operating characteristic curve (AUC) of 0.96 on validation set from the same region, but decreased to an AUC of 0.84 on test set from a different country. Classifier 2 demonstrated the most generalizable performance with an AUC of 0.96 on validation set from the same region and AUC of 0.95 on test set from a different country. Classifier 3 demonstrated the poorest performance, with an AUC of 0.83 on datasets from the same region and an AUC of 0.64 on datasets from a different country.
“Our classifiers achieved excellent performance and generalizability in the detection of PACD and moderate success in stratifying the different stages within PACD,” according to the researchers. “An automated image processing algorithm for a non-contact, effective clinical exam of the anterior chamber angle can empower population-based screening of PACD towards preventing irreversible vision loss.”
Study limitations include that the cohort was predominantly Chinese, failure to assess the performance of other convolutional neural network vision engines such as VGGNet or ResNet, and a need for more explainable classifiers, which can help address potential biases in these convolutional neural network classifiers, establish trust in the results and improve classifier performance.
Shan J, Li Z, Ma P, et al. Deep learning classification of angle closure based on anterior segment optical coherence tomography. Ophthalmol Glaucoma. Published online July 16, 2023. doi:10.1016/j.ogla.2023.06.011