Objective cataract detection and grading with deep learning based on OCT densitometry

Abstract

Purpose : To develop a deep convolutional neural network (DCNN) to robustly detect and grade nuclear cataract based on densitometry images obtained using deep-range anterior segment optical coherence tomography (AS-OCT).

Methods : A total of 1,309 eyes of 778 participants who underwent deep-range AS-OCT examinations were included. The tomographic images of the crystalline lenses were obtained using deep-range AS-OCT and the mean densities of the nuclei were evaluated. Lens opacities on slit-lamp images were graded according to the LOCS III system. AS-OCT images were by three ophthalmologists based on the densitometry and LOCS III scale. In total 1,309 high quality center OCT images were selected from 778 patients, in which 1074 images (320, 342, 256, and 265 images corresponding to normal, mild, moderate, and severe cases) were used to develop two DCNN models for cataract: End-to-end DCNN model and DCNN with a random forest classifier (DCNN+RF).A total of 235 images that were collected independently were used to test the DCNN model. Gradient Class Activation Map (GradCAM) were used to visualize the outcome and to evaluate the clinical relevance.
The area under the receiver operating characteristic curve (AUC), precision, recall, and R²metrics were used to evaluate the accuracy.

Results : The mean nuclear density was well correlated with the LOCS III NO scores (R²=0.73). The End-to-end DCNN had a better performance compared to DCNN+RF and detected nuclear cataract with the overall AUC of 0.97. The precision and recall were 0.88 and 0.86 respectively. The GradCAM map visualization indicated that our model made the prediction based on the features from the lens nucleus regions.

Conclusions : The OCT image based deep learning models precisely detect and grade the nuclear cataract. It has the potential to provide clinicians with an objective and robust detection nuclear cataract and a means for monitoring the cataract progression.

This is a 2021 Imaging in the Eye Conference abstract.

 

Figure 1. Scatter plots of the LOCS III score and nuclear densitometry (left), relationship between our dataset labeling and LOCS III score (middle), AUROC of two DCNN models (right).

 

Figure 2. GradCAM was plotted to visualize the attention of the last convolution layer.