IMAGE EVALUATION OF ARTIFICIAL INTELLIGENCE–SUPPORTED OPTICAL COHERENCE TOMOGRAPHY ANGIOGRAPHY IMAGING USING OCT-A1 DEVICE IN DIABETIC RETINOPATHY. Issue 8 (August 2021)

• Record Type: Journal Article
• Authors: Kawai, Kentaro; Uji, Akihito; Murakami, Tomoaki; Kadomoto, Shin; Oritani, Yasuyuki; Dodo, Yoko; Muraoka, Yuki; Akagi, Tadamichi; Miyata, Manabu; Tsujikawa, Akitaka
• Is Part Of: Retina. Volume 41:Issue 8(2021)
• Publication Date: 2021-08
• Subjects: artificial intelligence -- deep learning -- diabetic retinopathy -- imaging -- optical coherence tomography angiography; Retina -- Diseases -- Periodicals; Retinal Diseases; Vitreous Body; 617.735
• DOI: 10.1097/IAE.0000000000003101
• Languages: English
• ISSNs: 0275-004X

Abstract

Abstract : Purpose: To investigate the effect of denoise processing by artificial intelligence (AI) on the optical coherence tomography angiography (OCTA) images in eyes with retinal lesions. Methods: Prospective, observational, cross-sectional study. Optical coherence tomography angiography imaging of a 3 × 3-mm area involving the lesions (neovascularization, intraretinal microvascular abnormality, and nonperfusion area) was performed five times using OCT-HS100 (Canon, Tokyo, Japan). We acquired AI-denoised OCTA images and averaging OCTA images generated from five cube scan data through built-in software. Main outcomes were image acquisition time and the subjective assessment by graders and quantitative measurements of original OCTA images, averaging OCTA images, and AI-denoised OCTA images. The parameters of quantitative measurements were contrast-to-noise ratio, vessel density, vessel length density, and fractal dimension. Results: We studied 56 eyes from 43 patients. The image acquisition times for the original, averaging, and AI-denoised images were 31.87 ± 12.02, 165.34 ± 41.91, and 34.37 ± 12.02 seconds, respectively. We found significant differences in vessel density, vessel length density, fractal dimension, and contrast-to-noise ratio ( P < 0.001) between original, averaging, and AI-denoised images. Both subjective and quantitative evaluations showed that AI-denoised OCTA images had less background noise and depicted vessels clearly. In AI-denoised images, the presence of fictional vessels was suspected in 2 of the 35 cases of nonperfusion area. Conclusion: Denoise processing by AI improved the image quality of OCTA in a shorter time and allowed more accurate quantitative evaluation. Abstract : Artificial intelligence–denoised optical coherence tomography angiography images of diabetic retinopathy showed an equal or greater image quality than that of the averaged images in a shorter imaging time.