Deep learning segmentation of endothelial cell images using an active learning paradigm with guided label corrections.

Purpose: To create Guided Correction Software for informed manual editing of automatically generated corneal endothelial cell (EC) segmentations and apply it to an active learning paradigm to analyze a diverse set of post-keratoplasty EC images. Approach: An original U-Net model trained on 130 manually labeled post-Descemet stripping automated endothelial keratoplasty (EK) images was applied to 841 post-Descemet membrane EK images generating "uncorrected" cell border segmentations. Segmentations were then manually edited using the Guided Correction Software to create corrected labels. This dataset was split into 741 training and 100 testing EC images. U-Net and DeepLabV3+ were trained on the EC images and the corresponding uncorrected and corrected labels. Model performance was evaluated in a cell-by-cell analysis. Evaluation metrics included the number of over-segmentations, under-segmentations, correctly identified new cells, and endothelial cell density (ECD). Results: Utilizing corrected segmentations for training U-Net and DeepLabV3+ improved their performance. The average number of over- and under-segmentations per image was reduced from 23 to 11 with the corrected training set. Predicted ECD values generated by networks trained on the corrected labels were not significantly different than the ground truth counterparts (p=0.02, paired t-test). These models also correctly segmented a larger percentage of newly identified cells. The proposed Guided Correction Software and semi-automated approach reduced the time to accurately segment EC images from 15 to 30 to 5 min, an ∼80% decrease compared to manual editing. Conclusions: Guided Correction Software can efficiently label new training data for improved deep learning performance and generalization between EC datasets.

Comparison of Donor Corneal Endothelial Cell Density Determined by Eye Banks and by a Central Image Analysis Reading Center Using the Same Image Analysis Method.

PURPOSE: The purpose of this study was to evaluate agreement between eye banks (EBs) and an image analysis reading center on endothelial cell density (ECD) determinations using the same image analysis method. METHODS: The Cornea Image Analysis Reading Center (CIARC) determined ECD with a single experienced analyst on EB-obtained central endothelial images from donors intended for keratoplasty from 2 eye banks, Eversight and Lions VisionGift, using the Konan center analysis method. The EBs performed ECD determination on their respective sets of images using the same analysis method with experienced eye bank technicians. RESULTS: The mean age of the 200 donors was 54 years (range 30-75 years). Seventy (35%) of the 200 patients were women, and 57 (29%) were diabetic. The mean ECD was 10 cells/mm2 greater by the EBs than by CIARC (P = 0.39), with 95% limits of agreement of [-304 to 323 cells/mm2]. The mean difference was not substantially changed when the difference between EBs and CIARC ECD was adjusted for sex, donor age, donor diabetes, CV, HEX, number of cells analyzed, and EBs as a random effect (estimated mean difference of 20 cells/mm2 after adjustment in a linear mixed model; P = 0.73). The EB-determined preoperative ECD was within 10% of the CIARC-determined ECD for 178 (89%) image sets, with 15 (8%) higher by >10% and 7 (3%) lower by >10%. CONCLUSIONS: Well-trained eye bank technicians achieve comparable results for ECD determination with an experienced image analyst from an image analysis reading center when the same image analysis method is used.

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

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

Comparative corneal endothelial cell toxicity of differing intracameral moxifloxacin doses after phacoemulsification.

PURPOSE: To determine whether intracameral moxifloxacin 500 µg is noninferior to 250 µg for central endothelial cell loss (ECL) after phacoemulsification. SETTING: Aravind Eye Care System. DESIGN: Prospective masked randomized study. METHODS: Eyes with bilateral nuclear cataracts, central endothelial cell density (ECD) of more than 2000 cells/mm, and ECD not differing between eyes by more than 200 cells/mm underwent phacoemulsification at least 14 days apart. Intraoperatively, the first eye was randomized to receive either a 500 or 250 µg dose of moxifloxacin intracamerally and received the other dose for the second-eye surgery. Postoperative course was monitored at 1 day, 1 week, 1 month, and 3 months. Preoperative and 30-day and 90-day postoperative central ECD was determined by a reading center for a masked analysis of ECL at 3 months postoperatively. RESULTS: Fifty eyes of 25 patients (aged 48 to 69 years) underwent uneventful surgery and had normal postoperative courses. The point estimate (PE) and 95% CI for the mean difference in % ECL between the 500 µg and 250 µg doses at 3 months postoperatively was 0.8% (-5.8%, 7.4%). Upon identifying and removing 2 outliers, noninferiority was proven with a mean difference of the PE, -2.2% (CI, -6.5%, 2.1%). CONCLUSIONS: Clinical and corneal endothelial cell were comparable in this study population for the 250 µg and 500 µg doses of intracameral moxifloxacin. Both doses were well tolerated clinically, supporting the use of the higher dose for improved antimicrobial coverage for the prevention of postoperative endophthalmitis.

Machine learning for segmenting cells in corneal endothelium images.

Images of the endothelial cell layer of the cornea can be used to evaluate corneal health. Quantitative biomarkers extracted from these images such as cell density, coefficient of variation of cell area, and cell hexagonality are commonly used to evaluate the status of the endothelium. Currently, fully-automated endothelial image analysis systems in use often give inaccurate results, while semi-automated methods, requiring trained image analysis readers to identify cells manually, are both challenging and time-consuming. We are investigating two deep learning methods to automatically segment cells in such images. We compare the performance of two deep neural networks, namely U-Net and SegNet. To train and test the classifiers, a dataset of 130 images was collected, with expert reader annotated cell borders in each image. We applied standard training and testing techniques to evaluate pixel-wise segmentation performance, and report corresponding metrics such as the Dice and Jaccard coefficients. Visual evaluation of results showed that most pixel-wise errors in the U-Net were rather non-consequential. Results from the U-Net approach are being applied to create endothelial cell segmentations and quantify important morphological measurements for evaluating cornea health.

Postoperative Endothelial Cell Density Is Associated with Late Endothelial Graft Failure after Descemet Stripping Automated Endothelial Keratoplasty.

PURPOSE: To determine whether preoperative endothelial cell density (ECD) and postoperative ECD after Descemet stripping automated endothelial keratoplasty (DSAEK) are associated with late endothelial graft failure (LEGF) in the Cornea Preservation Time Study (CPTS). DESIGN: Cohort study within a multicenter, randomized clinical trial. PARTICIPANTS: A total of 1007 individuals (1223 study eyes), mean age 70 years, undergoing DSAEK for Fuchs' dystrophy (94% of eyes) or pseudophakic or aphakic corneal edema (PACE) (6% of eyes) and followed for up to 5 years. METHODS: Central ECD was determined by a central image analysis reading center. Preoperative ECD was determined for 1209 eyes that did not fail and 14 eyes that experienced LEGF. The ECD at 6 and 12 months after DSAEK, the change in ECD from preoperative to 6 and 12 months, surgeon-reported operative complications, and postoperative graft dislocation were investigated for an association with LEGFs unrelated to other postoperative events. Univariable and multivariable Cox proportional hazards regression models were used to assess associations. MAIN OUTCOME MEASURES: Late endothelial graft failure and its associations with pre- and postoperative ECD and operative complications. RESULTS: The cumulative probability of LEGF was 1.3% (95% confidence interval [CI], 0.8%-2.4%). Median (interquartile range [IQR]) preoperative ECDs were similar for eyes with LEGF (2523; 2367-3161) cells/mm2) and eyes without failure (2727; 2508-2973) cells/mm2) (P = 0.34). The ECD at 6 months was associated with LEGF (P < 0.001) in time-to-event analyses, whereas preoperative ECD was not (P = 0.55). The cumulative incidence (95% CI) of LEGF was 6.5% (3.0%, 14.0%) for 97 grafts with a 6-month ECD less than 1200 cells/mm2, 0.3% (0.0%, 2.4%) for 310 grafts with a 6-month ECD between 1200 and 2000 cells/mm2, and 0.6% (0.1%, 2.7%) for 589 grafts with a 6-month ECD greater than 2000 cells/mm2. In multivariable analyses, ECD at 6 months and operative complications were both associated with LEGF (P = 0.002 and P = 0.01, respectively), whereas graft dislocation was not (P = 0.61). CONCLUSIONS: In eyes undergoing DSAEK, preoperative ECD is unrelated to LEGF, whereas lower ECD at 6 months is associated with LEGF. Early endothelial cell loss after DSAEK and intraoperative complications should be minimized to improve graft survival.

Corneal Endothelial Cell Density and Morphology After Phacoemulsification in Patients With Primary Open-Angle Glaucoma and Cataracts: 2-Year Results of a Randomized Multicenter Trial.

PURPOSE: To evaluate corneal endothelial cell density (ECD) and morphology 2 years after phacoemulsification in subjects from the COMPASS trial (ClinicalTrials.gov, NCT01085357) who had mild-to-moderate primary open-angle glaucoma and visually significant cataracts. METHODS: The central corneal endothelium was evaluated by serial specular microscopy at 0 to 24 months. ECD, coefficient of variation, and percentage of hexagonal cells were evaluated by a central image analysis reading center and central corneal thickness (CCT) was evaluated by ultrasound pachymetry. RESULTS: Of 131 subjects who underwent routine phacoemulsification, analyzable endothelial images at 24 months were available for 126 subjects (96.2%). Mean ± SD central ECD at baseline was 2453 ± 359 cells/mm, decreasing by 10% ± 14% to 2195 ± 517 cells/mm at 3 months (P < 0.001) but stabilizing thereafter with mean endothelial cell loss (ECL) from baseline to 24 months of 9% ± 13% (P < 0.001). Twelve (9.5%) and 10 (7.9%) subjects experienced >30% ECL at 12 and 24 months, respectively. Neither coefficient of variation nor percentage of hexagonal cells changed significantly from baseline at any time point. Mean CCT was similar at baseline (550 ± 35 µm) and at 12 months (551 ± 37 µm) and 24 months (555 ± 35 µm). Age was significantly associated with ECL after cataract surgery (P = 0.02), but baseline intraocular pressure, number of glaucoma medications, and CCT were not. Similar results were observed in patients who underwent CyPass micro-stent implantation accompanying phacoemulsification. CONCLUSIONS: Phacoemulsification in eyes with mild-to-moderate primary open-angle glaucoma results in early ECL, with ECD stabilizing after 3 months and no effect on other endothelial stress markers up to 2 years postoperatively.

Comparison of Donor Cornea Endothelial Cell Density Determined by Eye Banks and by a Central Reading Center in the Cornea Preservation Time Study.

PURPOSE: To evaluate agreement between eye banks (EBs) and a reading center on endothelial cell density (ECD) determinations in the Cornea Preservation Time Study. METHODS: The Cornea Image Analysis Reading Center (CIARC) performed variable frame image analysis on EB-obtained-preoperative central endothelial images (after lamellar dissection for Descemet stripping automated endothelial keratoplasty by the EBs or before shipping, if surgeon prepared) to determine ECD. The EBs performed their usual method of ECD determination. The CIARC and EBs also provided ECD determinations from screening central endothelial images taken by the EBs during donor evaluation. Two independent masked CIARC readers determined ECD with measurements averaged. RESULTS: The mean preoperative ECD was 15 cells/mm greater by the EBs than by CIARC (N = 1286, P < 0.001) with 95% limits of agreement of (-644, 675 cells/mm). The limits of agreement in preoperative ECD were wider in the After-Lamellar-Dissection Group (-687, 683 cells/mm) than in the Before Shipping Group [(-505, 633 cells/mm); P = 0.03]. The EBs-determined preoperative ECD was within 10% of the CIARC-determined ECD for 886 (69%) image sets, with 236 (18%) higher by >10% and 164 (13%) lower by >10%. Excellent agreement appeared between the EBs and CIARC when 100-300 cells could be analyzed in contrast to <100 cells (SD = 308 cells/mm vs. SD = 603 cells/mm; P < 0.001). CONCLUSIONS: The mean ECD by the EBs and CIARC were similar, but there was considerable variability between determinations for individual corneas. Agreement improved between the 2 measurements when more than 100 cells were able to be analyzed.

Donor, Recipient, and Operative Factors Associated With Increased Endothelial Cell Loss in the Cornea Preservation Time Study.

Importance: Determining factors associated with endothelial cell loss after Descemet stripping automated endothelial keratoplasty (DSAEK) could improve long-term graft survival. Objective: To evaluate the associations of donor, recipient, and operative factors with endothelial cell density (ECD) 3 years after DSAEK in the Cornea Preservation Time Study. Design, Setting, and Participants: This cohort study was a secondary analysis of data collected in a multicenter, double-masked, randomized clinical trial. Forty US clinical sites with 70 surgeons participated, with donor corneas provided by 23 US eye banks. Individuals undergoing DSAEK for Fuchs dystrophy or pseudophakic/aphakic corneal edema were included. Interventions: The DSAEK procedure, with random assignment of a donor cornea with a preservation time of 0 to 7 days or 8 to 14 days. Main Outcomes and Measures: Endothelial cell density at 3 years as determined by a reading center from eye bank and clinical specular or confocal central endothelial images. Results: The study included 1090 participants (median age, 70 years) with 1330 affected eyes (240 bilateral cases [22.0%]), who underwent DSAEK for Fuchs dystrophy (1255 eyes [94.4%]) or pseudophakic/aphakic corneal edema (PACE) (75 eyes [5.6%]). Of these, 801 eyes (60.2%) belonged to women and 1207 (90.8%) to white individuals. A total of 749 participants (913 eyes; 164 [21.9%] bilateral cases) had functioning grafts with acceptable endothelial images preoperatively and at 3 years postoperatively and were included in this analysis. Factors associated with a lower ECD at 3 years (estimated effect with 99% CI) in the final multivariable model included donors with diabetes (-103 [-196 to -9] cells/mm2), lower screening ECD (-234 [-331 to -137] per 500 cells/mm2), recipient diagnosis of PACE (-257 [-483 to -31] in cells/mm2), and operative complications (-324 [-516 to -133] in cells/mm2). Endothelial cell loss (ECL) from a preoperative measurement to a 3-year postoperative measurement was 47% (99% CI, 42%-52%) for participants receiving tissue from donors with diabetes vs 43% (99% CI, 39%-48%) without diabetes; it was 53% (99% CI, 44%-62%) for participants diagnosed with PACE vs 44% (99% CI, 39%-49%) for those diagnosed with Fuchs dystrophy, and 55% (99% CI, 48%-63%) in participants who experienced operative complications vs 44% (99% CI, 39%-48%) in those who did not. No other donor, recipient, or operative factors were significantly associated with 3-year ECD. Conclusions and Relevance: Donor diabetes, lower screening ECD, a PACE diagnosis in the recipient, and operative complications were associated with lower ECD at 3 years after DSAEK surgery and may be associated with long-term graft success. While causation cannot be inferred, further studies on the association of donor diabetes and PACE in recipients with lower 3-year ECD warrant further study.

Effects of Repeated Intravitreal Aflibercept Injection on the Corneal Endothelium in Patients With Age-Related Macular Degeneration: Outcomes From the RE-VIEW Study.

PURPOSE: The effects of repeated intravitreal aflibercept injection (IAI) on the corneal endothelium were studied in patients with unilateral neovascular age-related macular degeneration. METHODS: RE-VIEW was a phase 4, open-label, single-arm, multicenter study. Patients received IAI every 8 weeks after 3 monthly doses. Slit-lamp biomicroscopy was performed at all study visits. The central corneal endothelial health was evaluated by specular microscopy in the treated versus untreated fellow eyes at baseline and weeks 24 and 52. RESULTS: No slit-lamp abnormalities were noted in 154 enrolled patients (eyes). Baseline versus 52-week mean (±SD) endothelial morphometric values (n = 118) for the treated versus untreated fellow eyes were respectively as follows: endothelial cell density was 2410 ± 364 versus 2388 ± 384 cells/mm at baseline and remained unchanged at 2401 ± 353 versus 2376 ± 364 cells/mm at 52 weeks (P = 0.87); the coefficient of variation was 33.5 ± 4.4% versus 34.0 ± 5.0% at baseline and remained unchanged at 34.2 ± 4.7% versus 34.1 ± 4.9% at 52 weeks (P = 0.18); the percentage of hexagonal cells was 59.5 ± 5.8% versus 59.6 ± 6.4% at baseline and remained unchanged at 59.5 ± 6.0% versus 59.5 ± 5.8% at 52 weeks (P = 0.96). CONCLUSIONS: Repeated IAI for 52 weeks had no apparent corneal endothelial toxicity noted on specular microscopy in patients treated for neovascular age-related macular degeneration.

Descemet's stripping automated endothelial keratoplasty: three-year graft and endothelial cell survival compared with penetrating keratoplasty.

PURPOSE: To assess 3-year outcomes of Descemet's stripping automated endothelial keratoplasty (DSAEK) in comparison with penetrating keratoplasty (PKP) from the Cornea Donor Study (CDS). DESIGN: Prospective, multicenter, nonrandomized clinical trial. PARTICIPANTS: A total of 173 subjects undergoing DSAEK for a moderate risk condition (principally Fuchs' dystrophy or pseudophakic corneal edema) compared with 1101 subjects undergoing PKP from the CDS. METHODS: The DSAEK procedures were performed by 2 experienced surgeons using the same donor and similar recipient criteria as for the CDS PKP procedures, performed by 68 surgeons. Graft success was assessed by Kaplan-Meier survival analysis. Central endothelial cell density (ECD) was determined from baseline donor and postoperative central endothelial images by the reading center used in the CDS Specular Microscopy Ancillary Study. MAIN OUTCOME MEASURES: Graft clarity and ECD. RESULTS: The donor and recipient demographics were comparable in the DSAEK and PKP groups, except that the proportion of Fuchs' dystrophy cases was higher in the DSAEK cohort. The 3-year survival rate did not differ significantly between DSAEK and PKP procedures performed for either Fuchs' dystrophy (96% for both; P = 0.81) or non-Fuchs' cases (86% vs. 84%, respectively; P = 0.41). Principal causes of graft failure or regraft within 3 years after DSAEK and PKP were immunologic graft rejection (0.6% vs. 3.1%), endothelial decompensation in the absence of documented rejection (1.7% vs 2.1%), unsatisfactory visual or refractive outcome (1.7% vs. 0.5%), and infection (0% vs. 1.1%), respectively. The 3-year predicted probability of a rejection episode was 9% with DSAEK versus 20% with PKP (P = 0.0005). The median 3-year cell loss for DSAEK and PKP was 46% and 51%, respectively (P = 0.33), in Fuchs' dystrophy cases and 59% and 61%, respectively (P = 0.70), in the non-Fuchs' cases. At 3 years, use of a smaller DSAEK insertion incision was associated with significantly higher cell loss (60% vs. 33% for 3.2- and 5.0-mm incisions, respectively; P = 0.0007), but not with a significant difference in graft survival (P = 0.45). CONCLUSIONS: The graft success rate and endothelial cell loss were comparable at 3 years for DSAEK and PKP procedures. A 5-mm DSAEK incision width was associated with significantly less cell loss than a 3.2-mm incision.

Effect of incision width on graft survival and endothelial cell loss after Descemet stripping automated endothelial keratoplasty.

PURPOSE: To assess the effect of incision width (5.0 and 3.2 mm) on graft survival and endothelial cell loss 6 months and 1 year after Descemet stripping automated endothelial keratoplasty (DSAEK). METHODS: One hundred sixty-seven subjects with endothelial decompensation from a moderate-risk condition (principally Fuchs dystrophy or pseudophakic corneal edema) underwent DSAEK by 2 experienced surgeons. The donor was folded over and inserted with single-point fixation forceps. This retrospective analysis assessed graft survival, complications, and endothelial cell loss, which was calculated from baseline donor and 6-month and 1-year postoperative central endothelial images evaluated by an independent specular microscopy reading center. RESULTS: No primary graft failures occurred in either group. One-year graft survival rates were comparable (98% vs 97%) in the 5.0- and 3.2-mm groups, respectively (P = 1.0). Complications included graft dislocation, graft rejection episodes, and elevated intraocular pressure and occurred at similar rates in both groups (P > or = 0.28). Pupillary block glaucoma did not occur in either group. Mean baseline donor endothelial cell density did not differ: 2782 cells per square millimeter in the 5.0-mm (n = 64) and 2784 cells per square millimeter in the 3.2-mm (n = 103) groups. Percent endothelial cell loss was 27% +/- 20% (n = 55) versus 40% +/- 22% (n = 71; 6 months) and 31% +/- 19% (n = 45) versus 44% +/- 22% (n = 62; 12 months) in the 5.0- and 3.2-mm incision groups, respectively (both P < 0.001). CONCLUSIONS: One year after DSAEK, overall graft success was comparable for the 2 groups; however, the 5.0-mm incision width resulted in substantially lower endothelial cell loss at 6 and 12 months.

Descemet's stripping automated endothelial keratoplasty outcomes compared with penetrating keratoplasty from the Cornea Donor Study.

PURPOSE: To assess outcomes 1 year after Descemet's stripping automated endothelial keratoplasty (DSAEK) in comparison with penetrating keratoplasty (PKP) from the Specular Microscopy Ancillary Study (SMAS) of the Cornea Donor Study. DESIGN: Multicenter, prospective, nonrandomized clinical trial. PARTICIPANTS: A total of 173 subjects undergoing DSAEK for a moderate risk condition (principally Fuchs' dystrophy or pseudophakic/aphakic corneal edema) compared with 410 subjects undergoing PKP from the SMAS who had clear grafts with at least 1 postoperative specular image within a 15-month follow-up period. METHODS: The DSAEK procedures were performed by 2 experienced surgeons per their individual techniques, using the same donor and similar recipient criteria as for the PKP procedures in the SMAS performed by 68 surgeons at 45 sites, with donors provided from 31 eye banks. Graft success and complications for the DSAEK group were assessed and compared with the SMAS group. Endothelial cell density (ECD) was determined from baseline donor, 6-month (range, 5-7 months), and 12-month (range, 9-15 months) postoperative central endothelial images by the same reading center used in the SMAS. MAIN OUTCOME MEASURES: Endothelial cell density and graft survival at 1 year. RESULTS: Although the DSAEK recipient group criteria were similar to the PKP group, Fuchs' dystrophy was more prevalent in the DSAEK group (85% vs. 64%) and pseudophakic corneal edema was less prevalent (13% vs. 32%, P<0.001). The regraft rate within 15 months was 2.3% (DSAEK group) and 1.3% (PKP group) (P = 0.50). Percent endothelial cell loss was 34+/-22% versus 11+/-20% (6 months) and 38+/-22% versus 20+/-23% (12 months) in the DSAEK and PKP groups, respectively (both P<0.001). Preoperative diagnosis affected endothelial cell loss over time; in the PKP group, the subjects with pseudophakic/aphakic corneal edema experienced significantly higher 12-month cell loss than the subjects with Fuchs' dystrophy (28% vs. 16%, P = 0.01), whereas in the DSAEK group, the 12-month cell loss was comparable for the 2 diagnoses (41% vs. 37%, P = 0.59). CONCLUSIONS: One year post-transplantation, overall graft success was comparable for DSAEK and PKP procedures and endothelial cell loss was higher with DSAEK.