Potential measurement errors due to image enlargement in optical coherence tomography imaging

Abstract

The effect of interpolation and super-resolution (SR) algorithms on quantitative and qualitative assessments of enlarged optical coherence tomography (OCT) images was investigated in this report. Spectral-domain OCT images from 30 eyes in 30 consecutive patients with diabetic macular edema (DME) and 20 healthy eyes in 20 consecutive volunteers were analyzed. Original image (OR) resolution was reduced by a factor of four. Images were then magnified by a factor of four with and without application of one of the following algorithms: bilinear (BL), bicubic (BC), Lanczos3 (LA), and SR. Differences in peak signal-to-noise ratio (PSNR), retinal nerve fiber layer (RNFL) thickness, photoreceptor layer status, and parallelism (reflects the complexity of photoreceptor layer alterations) were analyzed in each image type. The order of PSNRs from highest to lowest was SR > LA > BC > BL > non-processed enlarged images (NONE). The PSNR was statistically different in all groups. The NONE, BC, and LA images resulted in significantly thicker RNFL measurements than the OR image. In eyes with DME, the photoreceptor layer, which was hardly identifiable in NONE images, became detectable with algorithm application. However, OCT photoreceptor parameters were still assessed as more undetectable than in OR images. Parallelism was not statistically different in OR and NONE images, but other image groups had significantly higher parallelism than OR images. Our results indicated that interpolation and SR algorithms increased OCT image resolution. However, qualitative and quantitative assessments were influenced by algorithm use. Additionally, each algorithm affected the assessments differently.

Fig 1. Preparation of optical coherence tomography images.

Spectral-domain optical coherence tomography images were minified by a factor of four. They were then magnified by a factor of four to enlarge pixel size. Various interpolation and super-resolution algorithms were applied. The recovered image shown utilized the bicubic interpolation algorithm.

Fig 2. Peak signal-to-noise ratio (PSNR) for each type of recovered optical coherence tomography (OCT) image.

Interpolation algorithms examined included the bilinear (BL), bicubic (BC), and Lanczos3 (LA). The super-resolution (SR) algorithm was also investigated. Evaluated OCT images were obtained from normal eyes and from eyes with diabetic macular edema (DME). The higher the PSNR, the closer the degraded image was to the original. The human eye can detect PSNR differences >0.2 dB. *indicates P < 0.05 (paired t-test followed by Bonferroni correction).

Fig 3. Retinal nerve fiber layer (RNFL) delineation in original and recovered spectral-domain optical coherence tomography images.

Images show from the foveal center to 1.5 mm toward the optic disc (horizontal scan) (A) Original image (OR). (B) An enlarged minified image with no additional processing (NONE). The RNFL edge is jagged and is hardly identifiable near the fovea. (C) An enlarged minified image with the bilinear (BL) interpolation algorithm applied. The RNFL edge is smoother than in the NONE image, but the overall image remains blurred. (D) An enlarged minified image with the bicubic (BC) interpolation algorithm applied. The RNFL is less jagged than in the BL image. (E) An enlarged minified image with the Lanczos3 (LA) interpolation algorithm applied. The RNFL edge is smoother than the BL or BC images, with the most remarkable improvement near the foveal pit. However, the image still has a blurred edge. (F) An enlarged minified image with a learning-based super resolution (SR) algorithm applied. The RNFL edge has the best demarcation of all the magnified images, allowing the fovea to be as easily identified as in the OR image. However, texture details and tiny structures (e.g., capillaries) visible in the OR image are not fully depicted.

Fig 4. Retinal nerve fiber layer (RNFL) thickness measurements in normal subjects using different image processing techniques.

(A) The RNFL thickness measurements 0.5 mm from the fovea. Images recovered without interpolation (NONE), with the bicubic (BC) interpolation algorithm, and with the Lanczos3 (LA) interpolation algorithm resulted in significantly higher measurements than original images (OR). Measurements made on OR images and those recovered with the bilinear (BL) interpolation algorithm or the super-resolution (SR) algorithm were not statistically different. On the other hand, BL and SR images resulted in significantly thinner RNFL thickness measurements than in NONE images. (B) The RNFL thickness measurements 1.5 mm from the fovea. The BC images resulted in significantly thicker measurements than in NONE images. (C) The RNFL thickness measurements 3.0 mm from the fovea. The BL images resulted in significantly thinner RNFL measurements than in the OR images. Similar results were obtained 1.5 mm and 3.0 mm from the fovea. *indicates P < 0.05, paired t-test followed by Bonferroni correction.

Fig 5. Photoreceptor layer visualization in spectral-domain optical coherence tomography images of eyes with diabetic macular edema.

Images show a 1.0 mm horizontal scan centered on the presumed fovea. (A) Original image (OR). Photoreceptor layer status was graded as having a complete external limiting membrane (ELM) line and a discontinuous inner segment ellipsoid line (ISe). (B) An enlarged minified image with no processing (NONE). Aliasing occurred and the ELM and ISe band, which were hardly identifiable, were graded as absent. (C) An enlarged minified image with the bilinear (BL) interpolation algorithm applied. Recovered visualization of the ELM and ISe was achieved. However, an unexpected connection of the ISe band and the ELM (arrowheads) occurred. This did not exist in the OR image. Both the ELM and ISe band were graded as discontinuous. (D) An enlarged minified image with the bicubic (BC) interpolation algorithm applied. Both the ELM and ISe band were graded as discontinuous. (E) An enlarged minified image with the Lanczos3 (LA) interpolation algorithm applied. The RNFL edge is smoother than in the BL and BC images, but the rest of the image does not greatly differ. Both the ELM and ISe band were graded as discontinuous. (F) An enlarged minified image with the learning-based super-resolution (SR) algorithm applied. Although the image has a well demarcated RNFL layer, photoreceptor layer details remain absent. An unexpected connection between the ISe band and ELM is visible (arrowheads).

Fig 6. Parallelism of the photoreceptor-retinal pigment epithelium complex in spectral domain optical coherence tomography images in eyes with diabetic macular edema.

Parallelism reflects image complexity, which is used for quantitative evaluation of photoreceptor layer structural changes. Parallelism ranges from 0 to 1, with a more complicated, destructive changes resulting in higher parallelism values. Images processed with higher grade algorithms (e.g., super-resolution [SR], Lanczos3 [LA]) tended to have higher levels of parallelism. Additionally, all images processed with enhancing algorithms had significantly higher parallelism than the original (OR) and unprocessed (NONE) images, which were not statistically different from each other. This indicates that photoreceptor layer images were less complicated in processed images than in OR images. BL = bilinear, BC = bicubic.*indicates P < 0.05, paired t-test followed by Bonferroni correction.