In summary, a novel automated segmentation algorithm has been developed and evaluated. The segmentation algorithm is mainly based on dual-scale gradient information that captures both global and local features. As indicated by the reported high accuracy, reproducibility, and fast execution speed, it holds high promise toward practical applications.
Regarding accuracy, similar quantitative validations versus manual segmentation are found in four previous macular OCT image segmentation studies [18
]. In these studies, comparable border differences between algorithm-versus-manual-segmentation and segmenter-versus-segmenter were reported, although the paper that was based on a statistical model did not specify the absolute difference values [19
]. The authors of the 3D segmentation papers presented overall unsigned border differences of 5.69 ± 2.41µm (fovea excluded) [18
], 6.1 ± 2.9µm [22
] and 5.75 ± 1.37µm [23
] between algorithm and manual segmentation in their three studies. Our results (3.39 ± 0.96µm) are considerably smaller than what has been reported. While our algorithm may have been performing better, several factors may have contributed to our smaller values, including machine type (time domain OCT or SD-OCT), axial resolution, image type and quality, and the boundaries quantified. The mean difference between algorithm and manual segmentation shown in our study was slightly larger than the mean between-segmenter difference. The fact that all segmenters were well trained and given the same guidelines [28
] and that all our test images span the horizontal meridian may contribute to this result.
Among the nine boundaries segmented in our paper, the GCL/IPL and ELM are not commonly found in literature. Several attempts at quantifying the combined GCL and IPL thickness have shown its potential application in the detection of diseases such as glaucoma and diabetic retinopathy [6
]. Our results technically demonstrate the feasibility of automated GCL/IPL and ELM segmentation and/or quantification, which may be beneficial in future clinical study. With current OCT single image quality, manual segmentation of the GCL/IPL and ELM boundaries is not yet available. The IS/OS and OS/RPE were not tested for accuracy due to the availability of manual segmentation data [28
]. While visually these two untested boundaries appear to be performing well, they too should be validated against manual segmentation in future work.
The algorithm aims to achieve high speed without degrading accuracy. Ideally, to achieve the highest degree of sensitivity and accuracy, there would be customized Canny edges for each individual boundary detection. However, to reduce processing time in our implementation, the ILM and IS/OS share the information from the same Canny edge detector and axial gradient map, as do the OS/RPE and BM/Choroid, and the NFL/GCL and IPL/INL. The GCL/IPL and INL/OPL can be calculated separately in relatively little time, since the prior information of neighboring boundaries narrows down the detection area, which is an advantage of sequential detection. In this way, about 45 seconds are needed to complete the segmentation of a full-size 3D volume. To further gain speed, the A-scan reduction technique has been implemented, thereby successfully reducing the execution time to 16 seconds per volume without notably degrading the accuracy or reproducibility.
Although the algorithm is primarily based on gradient information, it still has the flexibility to add any other useful information to the final optimization because its post-processing is based on a convenient graph framework. Currently, only node cost is assigned. But link cost, usually assigned as smoothing terms, can also be used. If link cost were to be assigned, the additional smoothing procedure in the current algorithm may no longer be needed. Additional flexibility rests in the possibility for different kernel sizes, Canny thresholds for the various boundaries, and weight factors of different terms in the cost function. Our settings for these parameters were designed experimentally and intuitively based on various OCT images including normal and diseased eyes with different scan types. In other words, the algorithm has been optimized for general use rather than only for the data sets in this work.
In the final processing step for 3D volumes, the layer height information from neighboring B-scans is utilized to additionally smooth the 2D segmentation results. In our approach, the 3D information is integrated after 2D segmentation. Our current algorithm has the advantage of utilizing 3D information without building a complex 3D segmentation model, but it loses the ability to fully utilize the 3D information. In the future, efficiently integrating 3D information prior to 2D segmentation may be beneficial.
So far, this algorithm has been extensively evaluated for macular OCT images of relatively normal scans. More work is needed to evaluate our approach on retinas affected with outer and inner retinal diseases. Our approach can also be extended to other scan modes as well, such as circumpapillary and optic nerve head images.