**Purpose.**:
To demonstrate a novel algorithm for macular hole (MH) segmentation and volumetric analysis.

**Methods.**:
A computer algorithm was developed for automated MH segmentation in spectral-domain optical coherence tomography (SD-OCT). Algorithm validation was performed by trained graders with performance characterized by absolute accuracy and intraclass correlation coefficient. A retrospective case series of 56 eyes of 55 patients with idiopathic MHs analyzed using the custom algorithm to measure MH volume, base area/diameter, top area/diameter, minimum diameter, and height-to-base diameter ratio. Five eyes were excluded due to poor signal quality (1), motion artifact (1), and failure of surgical closure (3) for a final cohort of 51 eyes. Preoperative MH measurements were correlated with clinical MH stage, baseline, and 6-month postoperative best-corrected Snellen visual acuity (BCVA).

**Results.**:
The algorithm achieved 96% absolute accuracy and an intraclass correlation of 0.994 compared to trained graders. In univariate analysis, MH volume, base area, base diameter, top area, top diameter, minimum diameter, and MH height were significantly correlated to baseline BCVA (*P* value from 0.0003–0.011). Volume, base area, base diameter, and height–to-base diameter ratio were significantly correlated to 6-month postoperative BCVA (*P* value from <0.0001–0.029). In multivariate analysis, only base area (*P* < 0.0001) and volume (*P* = 0.0028) were significant predictors of 6-month postoperative BCVA.

**Conclusions.**:
The computerized segmentation algorithm enables rapid volumetric analysis of MH geometry and correlates with baseline and postoperative visual function. Further research is needed to better understand the algorithm's role in prognostication and clinical management.

^{ 1–4 }As such, the progression of MHs can be described by vitreoretinal forces that lead to a series of evolving anatomical defects. Advances in vitreoretinal surgery have contributed to successful postoperative restoration of foveal microstructure and restoration of visual function in most patients.

^{ 5–7 }Spectral-domain optical coherence tomography (SD-OCT) is a noncontact, noninvasive, laser interferometry technique that generates high-resolution, in vivo, cross-sectional images and is used in diagnosing numerous macular diseases. The application of SD-OCT imaging to MHs augments clinical staging by enabling visualization of the foveal and vitreous microstructure and tractional relationships and by calculating size measurements of hole architecture. SD-OCT allows for manual quantification of the width and height of the MH and identifies perifoveal cystoid edema, vitreomacular traction, and perifoveal inner segment and outer segment (IS/OS) integrity, and it allows clear discrimination between lamellar and pseudo-MHs.

^{ 8 }

^{ 9 }The ratio of the MH height-to-base diameter ratio (deemed the MH index, MHI) correlated with postsurgical visual function.

^{ 10 }The integrity of the perifoveal IS/OS has been associated with better postoperative visual outcome.

^{ 11 }

**Figure 1.**

**Figure 1.**

**Figure 2.**

**Figure 2.**

^{12,13}Briefly, an undirected graph

*G*= <

*V*,

*E*> is established by a set of vertices

*V*and edges

*E*. Consider the set of data elements IP representing image pixels and an 8-connected neighborhood system

*N*. For every element

*p*∈ IP, there exists an edge with nonnegative weight

*w*that connects

_{e}*q*∈

*N*. In addition, there are two vertices named the source

*S*and sink

*T*, each with edges that connect to every IP. A graph cut

*C*is a cut through the edges such that the induced graph

*G*= <

*V*,

*E*\

*C*> bipartitions the graph and separates the source and sink vertices. To use graph cuts to optimally segment a boundary, image intensity is encoded by edge weight to the sink and source vertices, while gradient is encoded by edge weight to neighboring nodes. A graph cut in the network computes the minimization of the energy function where

*R*(

*A*) specifies the intensity cost,

*B*(

*A*) specifies the boundary cost, and

*λ*is the intensity to boundary cost ratio. In our implementation, we define where

*I*is pixel intensity,

*C*and

_{F}*C*are finite foreground and background costs,

_{B}*σ*is a noise attenuation factor, and

*γ*is a nonlinear intensity scaling factor. Foreground seed pixels are established by thresholding image intensity in a truncated foveal region (Fig. 2C). The resulting seed markers are morphologically eroded to 25% of original area. For subsequent frames, the segmentation result from the adjacent frame is eroded to 25% of original area. Background seed pixels are established by selecting pixels a fixed distance from the centroid of foreground pixels.

*p*to the adjacent graph cut and

*α*is a reweight scaling factor. The graph cut is the set of edges that optimally segregates the hyporeflective MH cavity from the surrounding retina (Fig. 2D).

**Figure 3.**

**Figure 3.**

^{ 3,4 }All patients underwent standard three port 23 or 25-gauge pars plana vitrectomy (PPV) with or without peeling of the internal limiting membrane (ILM) assisted by indocyanine green staining, gas tamponade with perfluoropropane or sulfur hexafluoride, and prone face positioning. The clinical records were reviewed for baseline best-corrected Snellen visual acuity (BCVA) and Gass classification stage, as well as BCVA and hole closure verified by SD-OCT at postoperative month 1, 3, and 6 visits.

*P*value of less than 0.05 was considered statistically significant.

**Figure 4.**

**Figure 4.**

**Table 1.**

**Table 1.**

Stage 1 | Stage 2 | Stage 3 | Stage 4 | P Value | |

Mean volume, mm^{3} | 0.026 ± 0.014 | 0.037 ± 0.017 | 0.089 ± 0.071 | 0.179 ± 0.130 | 0.0005 |

Mean base area, mm^{2} | 0.147 ± 0.082 | 0.101 ± 0.088 | 0.364 ± 0.288 | 0.837 ± 0.668 | 0.0004 |

Mean base diameter, μm | 511 ± 145 | 416 ± 185 | 743 ± 285 | 993 ± 476 | 0.001 |

Mean top area, mm^{2} | 0.189 ± 0.168 | 0.285 ± 0.194 | 0.345 ± 0.185 | 0.511 ± 0.066 | 0.02 |

Mean top diameter, μm | 738 ± 229 | 752 ± 254 | 789 ± 189 | 934 ± 246 | 0.13 |

Mean minimum diameter, μm | N/A | 221 ± 136 | 323 ± 134 | 411 ± 172 | 0.0026 |

Mean height, μm | 189 ± 62 | 199 ± 79 | 330 ± 110 | 401 ± 105 | 0.0001 |

Macular hole index | 0.370 ± 0.049 | 0.516 ± 0.148 | 0.468 ± 0.127 | 0.515 ± 0.379 | 0.75 |

*P*value from 0.0003–0.011) with a correlation of larger measurement to worse visual acuity. The MHI did not correlate with baseline BCVA (

*P*= 0.163). When comparing the baseline MH measurements to the 6-month postoperative BCVA, hole volume, base area, base diameter, and MHI were highly correlated (

*P*value from <0.0001–0.029). There was no significant correlation of 6-month postoperative BCVA with top area (

*P*= 0.44), top diameter (

*P*= 0.58), minimum diameter (

*P*= 0.061), and height (

*P*= 0.80). The multiple regression model correlated age, baseline BCVA, and MH measurements (except base diameter and top diameter because they are redundant to base area and top area) to 6-month postoperative visual outcome. Only base area (

*P*< 0.0001) and volume (

*P*= 0.0028) were statistically significant predictors of 6-month postoperative BCVA (Table 3).

**Table 2.**

**Table 2.**

Spearman Correlation to Baseline BCVA | Spearman Correlation to 6-Month Postoperative BCVA | |

Volume | 0.488 (0.0003) | 0.306 (0.029) |

Base area | 0.443 (0.0011) | 0.470 (0.0005) |

Base diameter | 0.353 (0.011) | 0.371 (0.0074) |

Top area | 0.365 (0.0084) | 0.110 (0.442) |

Top diameter | 0.384 (0.0055) | 0.079 (0.580) |

Minimum diameter | 0.393 (0.0043) | 0.264 (0.061) |

Height | 0.369 (0.0077) | 0.248 (0.80) |

Macular hole index | −0.198 (0.163) | −0.380 (0.0059) |

**Table 3.**

**Table 3.**

P Value | |

Age, y | 0.52 |

Baseline BCVA, logMAR | 0.26 |

Volume, mm^{3} | 0.0028 |

Base area, mm^{2} | <0.0001 |

Top area, mm^{2} | 0.61 |

Height, μm | 0.97 |

Minimum diameter | 0.82 |

Macular hole index | 0.97 |

^{3}). The third eye was a chronic MH; although smaller, it had no macular edema and appeared to be nearly “flat open” at time of surgical intervention. With only three eyes in this group, statistical analysis was not performed on features associated with closure success.

^{ 3,4,8 }The staging criteria is generally associated with greater size of the MH cavity. MH volume, top area, top diameter, minimum diameter, and height increase with increasing stage. This trend is plausible given that volume, top area, and height are geometrically interrelated and correlate with a larger MH. Minimum diameter, measured as the minimum width in the OCT cross-section with greatest base diameter, can be considered the closest correlate to MH size in the Gass classification and was also seen to increase with increasing stage. Base area and base diameter also increase with increasing stage; however, base area and base diameter are larger in the stage 1 group than the stage 2 group. This may be due to the small sample size in the stage 1 group (

*n*= 3). Interestingly, the MHI was largest in the stage 2 group and stage 4 group.

^{ 11–14 }In this series, three eyes failed to close following surgical repair. Not surprisingly, a trend towards larger MH volume was seen. With only three eyes, the sample size was too small for statistical analysis. We are currently examining the algorithm's predictive value not only for surgical success (e.g., final closure) but also for rate of hole closure using perioperative OCT. If the high-resolution, 3D data set is able to provide information on hole closure rate, this tool may have significant utility in surgical decision-making (e.g., type of tamponade, duration of positioning).

^{ 15 }Multiple linear regression of the MH measurements adjusted for age and baseline BCVA demonstrated that only MH volume and base area were significant predictors of postoperative visual acuity. Previous studies have shown an association of the MHI to postoperative visual outcome.

^{ 10,11 }While we found a significant association in the univariate correlation, MHI was not a significant predictor of vision in the multiple regression model. We hypothesize that the lack of association is because the regression model included other parameters more strongly associated with postoperative visual acuity.

^{ 16–18 }In addition, SD-OCT studies examining the predictive potential of photoreceptor IS/OS measurements to visual outcome have been done.

^{ 11,19,20 }Our findings demonstrate that evaluating and including the noncentral OCT cross-sections of a MH and 3D MH morphology holds prognostic value, but more research is needed to demonstrate the comparative strength of its prognostication versus standard linear measurements.

*Eye (Lond)*. 2008; 22: 1301–1309. [CrossRef] [PubMed]

*Ophthalmology*. 1988; 95: 917–924. [CrossRef] [PubMed]

*Arch Ophthalmol*. 1988; 106: 629–639. [CrossRef] [PubMed]

*Am J Ophthalmol*. 1995; 119: 752–759. [CrossRef] [PubMed]

*Arch Ophthalmol*. 1997; 115: 11–21. [CrossRef] [PubMed]

*Ophthalmology*. 1999; 106: 1392–1397. [CrossRef] [PubMed]

*Am J Ophthalmol*. 2001; 132: 363–368. [CrossRef] [PubMed]

*Semin Ophthalmol*. 2003; 18: 58–66. [CrossRef] [PubMed]

*Arch Ophthalmol*. 2002; 120: 29–35. [CrossRef] [PubMed]

*Am J Ophthalmol*. 2004; 138: 709–716. [CrossRef] [PubMed]

*Invest Ophthalmol Vis Sci*. 2010; 51: 1651–1658. [CrossRef] [PubMed]

*Proceedings Eighth IEEE International Conference on Computer Vision 2001*. Los Alamitos, CA: IEEE Computer Society; 2001: 741.

*IEEE Trans Pattern Anal Mach Intell*. 2001; 23: 1222–1239. [CrossRef]

*Am J Ophthalmol*. 1999; 128: 655–657. [CrossRef] [PubMed]

*Br J Ophthalmol*. 2002; 86: 390–393. [CrossRef] [PubMed]

*Am J Ophthalmol*. 1998; 125: 204–213. [CrossRef] [PubMed]

*Br J Ophthalmol*. 2001; 85: 96–98. [CrossRef] [PubMed]

*Am J Ophthalmol*. 2002; 134: 667–674. [CrossRef] [PubMed]

*Am J Ophthalmol*. 2011; 152: 283–290. [CrossRef] [PubMed]

*Ophthalmologica*. 2011; 225: 47–54. [CrossRef] [PubMed]