A deep learning-based pachychoroid index based on choroidal image patterns of central serous chorioretinopathy using...

The study aimed to develop and evaluate a deep learning-based program called the Hokkaido University pachychoroid index (HUPI) from choroidal imaging patterns of enhanced-depth-imaging optical coherence tomography (EDI-OCT) in central serous chorioretinopathy (CSC), to function as the representative indicator of pachychoroid spectrum diseases.

A single center, retrospective, observational study.

A modified LeNet was trained and validated using choroidal images (128 x 128 pixels) extracted from EDI-OCT images of representative 37 CSC and 40 normal eyes. The HUPI was then generated from the algorithm, ranging between 0 and 1 (1: the most pachychoroid-like feature). Choroidal images were used to prevent learning of SRF information. To evaluate discriminative ability, the area under the curve (AUC) of HUPI was compared to conventional choroidal indices using 69 acute CSC and 100 normal eyes. The HUPI of 69 acute CSC eyes, 35 convalescent CSC eyes, 86 fellow eyes of 104 patients diagnosed with CSC, and 100 normal eyes were calculated for comparison.

HUPI showed a high ability to discriminate between acute CSC and normal eyes (AUC = 0.86), followed by the luminal to total choroidal area ratio (0.77), central choroidal thickness (0.76) and total choroidal area (0.75). The mean values of HUPI for normal, acute CSC, convalescent CSC, and CSC fellow eyes were 0.29±0.32, 0.77±0.28, 0.70±0.28 and 0.61±0.35, respectively. HUPI cut-off threshold was 0.66 with a sensitivity of 0.84 and a specificity of 0.77.

HUPI, a quantitative index of imaging patterns of pachychoroid, demonstrated a high ability to discriminate pachychoroid from normal eyes.