(C) PLOS One [1]. This unaltered content originally appeared in journals.plosone.org.
Licensed under Creative Commons Attribution (CC BY) license.
url:https://journals.plos.org/plosone/s/licenses-and-copyright

------------



OCTAVA: An open-source toolbox for quantitative analysis of optical coherence tomography angiography images

['Gavrielle R. Untracht', 'Optical Biomedical Engineering Laboratory', 'School Of Electrical', 'Electronic', 'Computer Engineering', 'The University Of Western Australia', 'Perth', 'Western Australia', 'Surrey Biophotonics', 'Advanced Technology Institute']

Date: 2022-01 

A comparison of the performance of each of the optimized segmentation algorithms for three images with different vascular network densities and SNRs can be seen in the top part of Fig 4 . In general, the built-in global thresholding algorithm within ImageJ does not accurately represent the complexity of the vascular network for in vivo OCTA MIP images of skin. It can be seen by visual inspection that the ImageJ global thresholding approach tends to underestimate the density of the vascular network and the resulting segmentation is poor. For images with high vessel density, such as the bottom row, all four of the remaining algorithms performed reasonably well; however, for images with lower VAD or poor SNR, ISODATA and adaptive thresholding tend to add additional structures within noisy but largely uniform regions of the image regardless of the kernel size specified. This means that these algorithms may have limited utility for images of skin but may still be useful for retinal images, which tend to have fewer uniform regions. The fuzzy thresholding and k-means algorithms replicate the vascular network in the binary mask over a larger variety of images. The five segmentation algorithms were also compared quantitatively using two metrics: VAD compared with manual calculation from the OCTA MIP image and network connectivity factor. The results of the quantitative comparison can be seen in the bottom of Fig 4 . The ISODATA and adaptive thresholding algorithms led to a high connectivity factor for all three images, which is not representative of the apparent differences in the images (higher CF does not necessarily correspond to better segmentation). The k-means and fuzzy thresholding algorithms performed similarly on both VAD and connectivity factor and led to a VAD closer to the manual calculation, indicative of good segmentation. Based upon this analysis performed over a large range of images, we have determined that the fuzzy thresholding algorithm provides the most accurate segmentation for our OCTA MIP images of skin and is applicable to a wider variety of images than the other algorithms. As a result, we have only implemented the fuzzy thresholding and adaptive thresholding algorithms in the final version of OCTAVA. While adaptive thresholding did not perform well with our images, it is used commonly in retinal OCTA; we included it only as a basis for comparison for users who may be more familiar with this algorithm.

3.2 Validation of vascular architecture metrics

To validate the accuracy of our generated metrics, we used two starting images: a conventional OCT en face image of a microfluidic device with known channel diameters; and a simulated OCT MIP image with hand-drawn structures mimicking a vascular network (Fig 5). The image of the microfluidic device was collected with a spectral domain OCT system described in [55]. The original OCT image depicted six small channels merging into one large channel. Although it conveys the advantage of known sizes, it was difficult to analyze this image since the pattern of the microfluidic device is not representative of a vascular network. Therefore, we synthesized a new image by copying and tiling the original image in a grid pattern to better mimic a network of vessels. For both images, the binary mask (Fig 5B and 5G) was generated using the fuzzy thresholding algorithm. Due to the bimodal distribution of vessels in the microfluidic image, the Frangi filter was not used; the performance of the Frangi filter is highly dependent on the range of vessel sizes within an image, so it is not possible to optimize the filter parameters for two distinct diameter ranges without distorting the image [59]. For the simulated OCTA MIP image, intensity variations were introduced along the simulated vessels so that the performance could be evaluated including the image enhancement step (without intensity variations, the Frangi filter would have no effect). The Frangi filter was used with a σ max value of 7 and was optimized using a procedure similar to that described in S1 File. The pixel size in each image (4 μm for the microfluidic image and 9.3 μm for the simulated OCTA MIP) determines the expected error bounds for diameter and length measurements.

PPT PowerPoint slide

PNG larger image

TIFF original image Download: Fig 5. Validation of image processing steps and metric results using images of a microfluidic device. (a-e) and a simulated OCTA MIP image (f-j). The composite image (a) is 4.32 mm × 4.32 mm with a pixel size of 4 μm. The binary mask (b) and overlay image (c) demonstrate the accuracy of the segmentation algorithm. The thickness map (d) demonstrates accurate measurement of the channel diameters throughout the image. The color bars in (d) and (i) indicate vessel diameter in μm. The green areas in the histogram indicate overlap between the blue and yellow bars. The same analysis was repeated for the simulated OCTA MIP image, which was assumed to be 10 mm × 10 mm with a pixel size of 9.3 μm. The overlay images (c, i) include a color-coded notation indicating different structures within the network: segments (yellow); branches (green); nodes (pink and blue circles); mesh regions (blue). (e) and (j) show the measured distribution of diameters in the microfluidic image and simulated OCTA MIP image, respectively. The larger channels in the microfluidic device are 300 μm and the smaller channels are 50 μm. https://doi.org/10.1371/journal.pone.0261052.g005

The segmentation was evaluated by plotting the intensity of the OCT image and the binary mask along a vertical or horizontal line. The results demonstrate that the segmentation step does not impact the apparent channel diameters in most cases, although the Frangi filter may increase the diameter of small, low-intensity vessels depending on the chosen value of σ max . This is further confirmed by the histogram of channel diameters for both images (Fig 5E–5J). As expected, the histogram of diameters for the microfluidic device indicates a bimodal distribution with peaks at the correct values of 50 μm and 300 μm representing the small and large channels. Measured diameters are typically within the error bounds determined from the pixel size with a few exceptions. For example, some of the channels in the microfluidic device are measured to have larger diameters than expected due to the functioning of the algorithm at junctions between the small and large channels, as can be seen from the overlay image (Fig 5C). Someone manually characterizing the image might identify a node at the endpoint of each of the 50-μm channels before any broadening or change in direction indicating the junction with the larger channel; the automated characterization often includes part of this transition region as part of the smaller channels and places the node closer to the 300-μm channel. This can be seen most clearly in the top and bottom channels of each row. As a result, the measured average diameter of the 50-μm channel is larger than it would be if the transition region were identified as an independent segment. For comparison, the nodes identified by OCTAVA were taken to be the segment endpoints for the manual characterization of the network. The thickness maps (Fig 5D–5I) demonstrate accurate diameter measurements throughout both images. The histograms of the lengths also match closely with the expected values (not shown).

The metrics generated by OCTAVA were compared with values calculated manually for both the microfluidic image and simulated OCTA MIP image (Table 2). The VAD was calculated manually using the mean intensity of the image as a global threshold. This is a valid approach for these images since there is a clear difference in intensity level between the network of channels and the background. Diameter and length measurements for the microfluidic images were measured manually in ImageJ in order to manually calculate VLD, mean and median diameter, and branchpoint density; for the simulated OCTA MIP image, exact diameter and length measurements were calculated based on a priori knowledge of the image structure. In all but one measurement, the relative difference between the automatic and manual metric calculations did not exceed 10%, although OCTAVA tends to overestimate the vessel diameter when compared with manual characterization based on the way it identifies nodes and vessel endpoints as described above. The automated method we use for measuring the vessel diameter (the local thickness algorithm in ImageJ) is commonly used for measuring the diameter of structures within an image (irrespective of the image content), however, it is difficult to replicate it exactly in manual analysis. This difficulty, in conjunction with the placement of nodes described above, and the limitation based on pixel size, accounts for the differences in the diameter measurements.

[END]

[1] Url: https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0261052

(C) Plos One. "Accelerating the publication of peer-reviewed science."
Licensed under Creative Commons Attribution (CC BY 4.0)
URL: https://creativecommons.org/licenses/by/4.0/


via Magical.Fish Gopher News Feeds:
gopher://magical.fish/1/feeds/news/plosone/