Comparison of Weka-assisted and conventional Niblack-based choroidal binarization in central serous chorioretinopathy

To compare a Trainable Weka Segmentation workflow (WEKA) with conventional Niblack-based methods for choroidal binarization in central serous chorioretinopathy (CSC). This was a retrospective observational comparative study. In a validation set of optical coherence tomography (OCT) B-scan images from 10 CSC eyes and 10 control eyes, segmentation quality was assessed by masked graders. In a separate cohort of 30 CSC eyes and 30 controls, the choroidal region was binarized using three methods: two conventional Niblack-based approaches (NB-A and NB-S) and a supervised WEKA classifier. Quantitative analyses included inter-method choroidal vascularity index (CVI) differences, pixel-level mismatch, and correlations between CVI and central choroidal thickness (CCT). In CSC eyes, WEKA achieved significantly higher qualitative segmentation scores than NB-A and NB-S (both p < 0.001) and produced higher CVI values (both p < 0.001). NB-A – and NB-S–based CVI values were significantly lower in CSC eyes than in controls (both p < 0.001), whereas WEKA-based CVI did not show this between-group reduction. Pixel-level mismatch rates involving WEKA (NB-A–WEKA and NB-S–WEKA) were significantly greater in CSC eyes than in controls (both p < 0.001). Pixel-level mismatch heatmaps showed inter-method disagreement within regions corresponding to dilated choroidal vessels in CSC eyes. Of the three methods, only WEKA-derived CVI was significantly correlated with CCT in both CSC ( r = 0.63, p < 0.001) and control eyes ( r = 0.59, p < 0.001). WEKA achieved higher qualitative segmentation scores and yielded substantially different CVI estimates compared with conventional Niblack-based methods. These findings suggest that binarization method selection can influence CVI assessment and interpretation in CSC. Not applicable.