A machine learning method for optical coherence tomography scan quality assessment

Abstract

Purpose : The reliability of automated analysis of optical coherence tomography (OCT) scans depends on the scan quality. Quality indicators in commercial instruments only provide an overall score and do not provide localized information. Here we demonstrated a quality map using a semi-supervised machine learning technique which can aid with identifying local areas of poor quality.

Methods : Our method first computes a set of feature maps using signal strength, signal-to-noise ratio (SNR), and contrast for individual or a group of neighboring A-scans of 580 6x6x2mm OCT volumes with good and poor quality (Fig 1A shows one volume). It then combines the feature maps into a joint probability map. Then, it generates groups of likelihood functions as inference models for the feature maps using a series of thresholds (0.1 to 0.5) on the joint probability map (Fig 1B shows one group). The quality map of a volume is generated using a posterior probability for each inference model (Fig 1C). The quality score for each B-scan is calculated by averaging a quality map along the fast B-scan. To select the best inference model, we used B-scans acquired using CIRRUS^TM HD-OCT 5000 (ZEISS, Dublin, CA) along with their corresponding B-scan quality scores and quality grading (performed by clinical experts) to compute receiver operating curves (ROC) (Fig 1D). Finally, the inference model with greatest area under the curve (AUC) was selected as the best inference model. A B-scan quality indicator threshold (0.3) was determined using a defined sensitivity (0.85) and specificity (0.9) on the ROC associated with the chosen inference model.

Results : Performance of the algorithm was evaluated using 481 B-scans of different quality but of the same scan type as the training data. Fig. 2 shows color-coded quality maps (red: highest quality, blue: lowest quality) and examples of OCT B-scans extracted from the volume. The sensitivity and specificity were found to be 0.89 and 0.84, respectively.

Conclusions : We demonstrated a new OCT scan quality assessment approach based on a semi-supervised machine learning method that yields a comprehensive quality map - an essential tool prior to OCT data analysis.

This is a 2020 Imaging in the Eye Conference abstract.

 

Fig 1. (A) A joint probability map from OCT volume; (B) Likelihood functions from feature maps; (C) Quality map from feature maps and the likelihood functions; (D) ROCs using three inference models.

 

Fig. 2: Examples of quality maps and corresponding B-scans.