Using Deep Learning for Assessing Image-Quality of 3D Macular Scans from Spectral-Domain Optical Coherence Tomography

Abstract

Purpose : Assessing image quality of optical coherence tomography (OCT) beforehand is very essential for subsequent retinal disease identification. We developed, validated, and tested deep-learning algorithms to assess the quality of macular scans obtained from two OCT devices.

Methods : This study was a retrospective analysis of OCT images obtained from the Chinese University of Hong Kong Eye Centre and Hong Kong Eye Hospital. The OCT images of different scanning protocols and devices (macular cube 512×128 B-scans from Cirrus OCT; and macular volume 512×31 B-scans, 1024×25 B-scans, and 1024×19 B-scans from Spectralis OCT) were included. The gradability of B-scan images and volume scans were labeled as gradable or ungradable. Gradable B-scan/volume scan was defined as absence of any artifact or; OCT artifacts affect less than 25% peripheral area or; fovea off centration presents in the central 50% area. Ungradable B-scan/volume scan was defined as OCT artifacts affect the central 50% area or fovea off centration is outside the central 50% area or fovea motion artifact. The ground truth was independently labeled by two masked well-trained graders. The discrepancy during the grading was resolved by adjudication by a senior grader. A total of 2,277 Cirrus scans and 33,633 Spectralis B-scans of 1,557 volume scans were divided into training (70%), validation (20%), test (10%), respectively. We developed, validated, and tested a 3D residual network (ResNet)-18 for Cirrus 3D cube scans, a dense convolutional network (DenseNet)-121 for Spectralis 2D B-scans, and a multiple-instance learning model for Spectralis 3D volume scans (Figure 1).

Results : In the primary validation of Cirrus cube scans, Spectralis B-scans, and Spectralis volume scans, the algorithms achieved the area under receiver operating characteristic curves (AUCs) of 0.930, 0.972, and 0.906 (Figure 2); sensitivities of 94.6%, 90.1%, and 86.5%; specificities of 83.3%, 94.2%, and 95.7%; and accuracies of 93.3%, 91.1%, and 87.2%, respectively.

Conclusions : The proposed deep learning algorithms achieved good performance, to distinguish gradable and ungradable OCT macular images. Incorporating with an artificial intelligence-based model, a volume-level quality indicator allows only gradable scans to be referred, which can smooth clinical operational flow for enhancing disease screening and diagnosis.

This abstract was presented at the 2022 ARVO Annual Meeting, held in Denver, CO, May 1-4, 2022, and virtually.

 

Figure 1

 

Figure 2