Automatic Detection of Choroidal Vascular Hyperpermeability Using Machine Learning Strategies

Abstract

INTRODUCTION: Our purpose was to develop an artificial intelligence system capable of accurately predict choroidal vascular hyperpermeability (CVH) using swept-source optical coherence tomography (SS-OCT).
METHODS: This was a retrospective observational study conducted in healthy and CSC patients. All cases underwent ultra-widefield (UWF) indocyanine green angiography (ICGA) using a confocal scanning laser ophthalmoscopy (SLO) device and widefield (WF) SS-OCT. To segment the choroid in each individual B-scan, an automatic segmentation model was created and its performance was evaluated. Thickness maps were generated. Then, to predict the presence of choroidal hyperpermeability, based on the choroidal thickness maps, we developed an automatic classification model. To evaluate this classification model a validation dataset was used, superimposing the generated thickness maps with ICGA scans.
RESULTS: A total of 26 eyes from 13 CSC patients and 14 eyes from 7 healthy patients were included. In our analysis, the automatic segmentation model achieved a precision of 94.72%. The automatic hyperpermeability classification model based on SS-OCT had a false-positive rate of 0%, but a false-negative rate of 18.75%.
CONCLUSION: In the presence of a high clinical suspicion of CVH, the automated classification model for hyperpermeability based on SS-OCT has proven to be a useful diagnostic method. However, due to the rate of false negatives, in cases where clinical suspicion remains high and the model yields a negative result, it may be worth considering performing an ICGA.