Abstract
Purpose :
The quality of optical coherence tomography angiography (OCTA) images will greatly impact their interpretability for doctors and the reliability of the downstream quantitative analysis, while the quality assessment is a subjective task that has a high demand on the prior experience. This study introduces a method to automatically generate an OCTA quality indicator using machine learning for operators' and doctors' references.
Methods :
OCTA data used in the study were captured by swept-source OCT devices (DRI OCT Triton; Topcon Corporation, Tokyo, Japan). The quality of OCTA superficial capillary plexuses (SCP) en face images was assessed separately by one grader from 1 (lowest) to 5 (highest), any score =1 is recommended for a retake and any score <=2 is highly possible to lead to errors in downstream quantification.
Twenty-eight (28) parameters that characterize features/artifacts observed in OCTA images related to image quality (e.g. OCT-related features: focus, folding/clipping, blink, motion, scan difficulty; OCTA-specific features: uneven illumination/shadow, motion artifact/microsaccades) were designed and represented by numerical values.
The proposed algorithm involves a two-layer stacked supervised model. The first layer consists of multiple weak classifiers (e.g. logistic regression, random forest tree, decision tree), while the second layer is a meta-model that determines the best way to combine the predictions of the first-layer models. The model was trained on the extracted feature parameters and human grades.
Results :
A total number of 209 OCTA data (including 3x3mm,4.5x4.5mm, and 6x6mm) from 21 subjects was captured. Among 21 subjects, 9 of them reported diabetic retinopathy. The stacked model demonstrated strong performance on Triton OCTA SCP enface images, with an area under the ROC curve (AUC) of 0.92. This represents a significant improvement over the single machine learning model, whose AUC was 0.83.
Conclusions :
The proposed stacked model was able to automatically assess the quality of OCTA scans with a higher AUC than the traditional single-model method. The stable performance in a small dataset shows its potential in applications for medical images. In this study, an objective and quantitative image quality indicator is provided for operators on retake decisions and indicates the quantification reliability for researchers and doctors.
This abstract was presented at the 2023 ARVO Annual Meeting, held in New Orleans, LA, April 23-27, 2023.