gms | German Medical Science

Purpose: Rhegmatogenous retinal detachment (RRD) is a sight-threatening emergency, where longer foveal detachment linked to worse visual outcomes. Fast classification of the duration of central vision loss is critical for timely surgical intervention but can be challenging. Sequential retinal changes on OCT offer valuable insights into foveal detachment, while duration is mostly estimated by relying onto subjective patient history. This study aims to evaluate a custom pre-trained AI model for recognizing OCT changes and associating them with fovea-involvement duration.

Setting/venue: This retrospective study was conducted at the University Medical Center Hamburg-Eppendorf’s Department of Ophthalmology, involving outpatients with RRD between 2019 and 2023.

Methods: This retrospective study was conducted at the University Medical Center Hamburg-Eppendorf’s Department of Ophthalmology, involving outpatients with RRD between 2019 and 2023, by whom an OCT was performed at presentation and the duration of central loss vision was clearly defined. This two-class classification task uses 2D OCT imaging to classify RDD. Class 1 includes physiological OCTs, macula-on RDD, and macula-off RDD lasting more than 3 days. Class 2 includes macula-off RDD lasting 3 days or less. A custom pre-trained convolutional neural network (CNN) was used to perform the classification. The model’s performance was evaluated using a 95% confidence interval (CI) calculated via bootstrapping, with the standard deviation (SD) included to represent the variability.

Results: Our modified pre-trained AI model was evaluated using One versus One ROC-AUC and precision-recall curves. The ROC-AUC showed significant performance, with an average AUC of 0.80 ± 0.02, highlighting excellent discrimination between classes. The precision-recall curve achieved an average precision of 0.82 ± 0.02 for detecting macula-off RDDs ≤3 days versus the differential group. These results confirm the model’s effectiveness in identifying clinically relevant OCT changes.

Conclusions: The presented deep learning model was able to significantly identify the macular involvement in cases of RRD, as well as to successfully determine if the foveal detachment had occurred more or less than 3 days before presentation, proving to be a useful and valuable tool facilitating the decision-making process of retinal specialists in their daily clinical setting aiding in the identification of RD severity and urgency.