Automated Image Quality Assessment for Spectralis SD-OCT Scans Utilizing Intrinsic Image Features and ML Confidence Score

Abstract

Purpose : To develop an automated Random Forest (RF) machine learning model for evaluating Spectralis (Heidelberg) SD-OCT image quality.

Methods : This analysis utilized 52 SD-OCT Spectralis SD-OCT scans for training. Each scan was reviewed by two expert graders on a slice-by-slice basis. Each slice was assigned to one of three quality categories (Good, Moderate, Poor), with Good being defined as an image with no quality defects, Moderate being a gradable image with noticeable quality defects, and poor being an ungradable image. Following initial assessment, 4 additional limited quality scans were introduced to provide greater representation of “Poor” quality slice examples. Image features including intensity mean, median, variance, skew, kurtosis, and homogeneity were extracted. In addition to these features, a set of machine learning layer segmentation confidence scores were included in the feature vector for each image.
The computed features were used to train a series of random forest classifiers. To assess the impact of class imbalance on feature training, excess good and moderate scans were filtered from the training folds until a desired class balance (2:2:1) was achieved. Mean AUC scores for each class were computed via One vs. Rest binarization and micro-averaging over 10 independent runs of 5-fold cross validation.

Results : Following initial training, micro-averaged AUC scores of 0.90 ± 0.03, 0.85 ± 0.05, and 0.82 ± 0.12 for Good, Moderate, and Poor, respectively were achieved. Using the expanded dataset with increased representation of poor quality scans, the performance of the classifier on was improved (Good: 0.91 ± 0.06, Moderate: 0.82 ± 0.07, Poor: 0.90 ± 0.06). Class balancing the training sets via a filtering algorithm and including machine learning confidence scores resulted in similar but slightly improved mean AUC metrics (Good: 0.93 ± 0.03, Moderate: 0.84 ± .07, Bad: 0.90 ± 0.05) with more balanced confusion matrices.

Conclusions : The random forest model enabled reliable automated classification of “Good” and “Poor” quality Spectralis OCT images with minimal cross-over between these two classes. This tool has the potential to act as an image quality-gating system by providing rapid feedback during clinical trial image acquisition and advanced image analysis programs.

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