Abstract
Purpose :
Excluding poor quality images is important both for clinicians to reliably assess the current status of the eye and for scientists to make valid inferences regarding biomarkers, group-differences or drug efficacy. However, manually reviewing images is time-consuming and is limited by reviewer biases. Here, we developed a machine learning classifier to provide a scalable and objective approach that can automatically detect poor quality images.
Methods :
An in-house database of 823 NIRAF human retinal images were manually graded as “bad” or “good” quality by two observers. For each image, image characteristics were captured using a set of 23 features (Figure 1d). This included no-reference (e.g., signal-to-noise ratio) and pseudo-reference metrics (e.g., mutual information) where the image metrics was compared to random uniform noise. A gradient boosting classifier was trained and tested using random 80-20% splits (500 “weak learners”, 0.01 learning rate). This process was repeated 250 times to build performance distributions to compare the model performance to random chance.
Results :
Manual observers rated ~28% (237) of the images as poor quality and ~72% (586) as good quality. The trained model correctly classified image quality at an f1-score performance (Figure 1ai) that was significantly above random chance (paired t-test: t(249)=101.59, p<0.001, d=9.07). On average, the model performance was high (Figure 1aii): f1-score=76% (SD=4.8), accuracy=87% (SD=2.5), precision=83% (SD=6.1), sensitivity=71% (SD=6.5%), specificity=94% (SD=2.3). Examining the model weights (Figure 1d) indicated that measures which capture the image’s information content and difference from random noise are most useful for detecting poor image quality.
Conclusions :
We demonstrate that machine learning can accurately classify human NIRAF retinal image quality. This approach could be applied in several contexts including improving quality during the real-time image acquisition or for triaging large numbers of images where distinct processing pipelines could be applied based on image quality.
This abstract was presented at the 2023 ARVO Annual Meeting, held in New Orleans, LA, April 23-27, 2023.