Abstract
Purpose :
To design and validate an innovative deep learning system that automates the detection of anomalies in retinal function from full-field electroretinogram (ERG) signals.
Methods :
A two-stage automated system comprising ERG preprocessing and deep learning classification was developed to analyze and identify anomalies in retinal function. For each patient, twelve ERG tracings were recorded under scotopic (0.01 cd.s/m2, 3.0 cd.s/m2, 3.0 cd.s/m2 OP, and 10.0 cd.s/m2) and photopic (3.0 cd.s/m2 cone and 3.0 cd.s/m2 flicker) conditions, capturing data for 220 milliseconds at a sampling rate of 1000 Hz from each eye. A deep learning (DL) classification model consisting of a 1D convolutional layer followed by three residual blocks, and a fully connected layer with a sigmoid activation function, was designed to classify the ERGs into two categories: “Normal” and “Abnormal.” An ERG dataset, which included 1,188 ERG tracings recorded from 99 patients (30 Normals, 69 Abnormals) along with diagnoses from electrophysiologists, was established for training and validating the DL model. Diagnoses for patients whose ERGs were considered Abnormal included cone dystrophy, macular dystrophy, and retinitis pigmentosa.
Results :
The deep learning model was trained on 70% of the ERG dataset (70 patients) for 100 epochs with a batch size of 16, using the Adam optimizer with default parameters and an initial learning rate of 0.0001. Subsequently, the trained model was evaluated on the remaining 30% of the dataset (29 patients) and yielded a weighted F1 score of 93.14% (91.67% for the “Normal” category and 94.12% for the “Abnormal” category).
Conclusions :
The proposed deep learning-assisted electroretinogram (ERG) analysis system featuring a 12-tracing measurement protocol demonstrates effectiveness in detecting anomalies in retinal function on full-field ERG. This system can be used as an effective tool to assist electrophysiologists and ophthalmologists by providing rapid automated assessments (to speed up interpretation) and decision support for ERG analysis.
This abstract was presented at the 2024 ARVO Annual Meeting, held in Seattle, WA, May 5-9, 2024.