Quantitative and qualitative evaluation of deep learning automatic segmentations of corneal endothelial cell images of reduced image quality obtained following cornea transplant

摘要: We are developing automated analysis of corneal-endothelial-cell-layer, specular microscopic images so as to determine quantitative biomarkers indicative corneal health following transplantation. Especially on these varying quality, commercial image systems can give inaccurate results, and manual methods very labor intensive. have developed a method automatically segment endothelial cells with process that included flattening, U-Net deep learning, postprocessing create individual cell segmentations. used 130 one type transplantation (Descemet stripping keratoplasty) expert-reader annotated borders. obtained good pixelwise segmentation performance (e.g., Dice coefficient  =  0.87  ±  0.17, Jaccard index  =  0.80  ±  0.18, across 10 folds). The segmented left unmarked by analysts sometimes differently than was split or two were merged). A clinically informative visual the held-out test set showed 92% within manually labeled regions acceptably that, compared segmentation, automation added 21% more correctly cells. speculate could reduce 15 30 min 3 5 min review editing.