Retinal Shape-Based Classification of Retinal Detachment and Posterior Vitreous Detachment Eyes

Abstract

Introduction: Retinal detachment is a sight-threatening emergency, with more than half of those affected suffering permanent visual impairment. A diagnostic test to identify eyes at risk before vision is threatened would enable exploration of prophylactic treatment. This report presents the use of irregularities in retinal shape, quantified from optical coherence tomography (OCT) images, as a biomarker for retinal detachment.

Methods: OCT images were taken from posterior and mid-peripheral retina of 264 individuals [97 after a posterior vitreous detachment (PVD), 99 after vitrectomy for retinal detachment and 68 after laser for a retinal tear]. Diagnoses were taken from history, examination and OCT. Retinal irregularity was quantified in the frequency domain, and the distribution of irregularity across the regions of the eye was explored to identify features exhibiting the greatest difference between retinal detachment and PVD eyes. Two of these features plus axial length were used to train a quadratic discriminant analysis classifier. Classifier performance was assessed by its sensitivity and specificity in identifying retinal detachment eyes and visualised with a receiver operating characteristic (ROC) curve.

Results: Validation set specificity was 84% (44/52 PVD eyes correctly labelled) and sensitivity 35% (23/64 retinal detachment eyes identified, p = 0.02). Area under the ROC curve was 0.75 (95% confidence intervals 0.58-0.85). Retinal detachment eyes were significantly more irregular than PVD eyes in the superior retina (0.70 mm versus 0.49 mm, p < 0.05) and supero-temporal retina (1.12 mm versus 0.80 mm, p < 0.05). Lower sensitivity (16/68, 24%) was seen for eyes with a retinal tear without detachment, that were intermediate in size between retinal detachment and PVD eyes. Axial length on its own was a poor classifier. Neither irregularity nor classification were affected by surgery for retinal detachment or the development of PVD.

Conclusions: The classifier identified 1/3 of retinal detachment eyes in this sample. In future work, these features can be evaluated as a test for retinal detachment prior to PVD.

Keywords: Fourier analysis; Machine learning; Optical coherence tomography; Retinal detachment; Vitreous detachment.

Fig. 1

Flow chart. The data set of retinal detachment (RD) and PVD eyes was split into training and testing sets. Shape features were extracted from the training set images (Fig. 2) and explored to find useful differences between retinal detachment and PVD eyes. These were combined with axial length to train a classifier. The classifier was tested with the validation set of retinal detachment and PVD eyes, and a separate validation set of retinal tear eyes. Results are presented in a confusion matrix in Table 3

Fig. 2

Shape analysis and feature extraction. The retinal shape (taken from the retinal pigment epithelial line) in the B scan (i), is plotted in (ii). The middle figure (iii) shows the residual left after subtraction of the best fit quadratic curve. Figure (iv) is the Fourier transform of (iii) and describes the retinal irregularity in the frequency domain. The 30 lowest frequency bins are magnified in (v). The difference between an individual B scan irregularity [dotted line in (vi)] and the average B scan irregularity [solid line in (vi)] was used for analysis (Supplementary information 1). Each vertical coloured bar represents the contribution of a range of frequencies (a bin) to the irregularity and is used for analysis of shape. Note the B scan image (i) is shown as presented in the clinic, and it is not uncommon for mid-peripheral retina to be as irregular as this. The true aspect ratio is 9 mm (wide) × 2 mm (deep)

Fig. 3

Receiver operating characteristic curve. For identification of retinal detachment eyes. Error bars representing 95% confidence intervals (CI) were generated from 5000 bootstrap replicas. Area under the curve = 0.75 (95% CI 0.58–0.85)

Fig. 4

Bland–Altman test–retest plots of shape metrics (17 regions, frequency bin moduli 2–20) from B scan irregularity spectra before and after PVD [blue points (i)], and before PVD and after vitrectomy for retinal detachment [red points (ii)]. The 95% limits of agreement are set by test–retest measurements from eight eyes with no change in status between examinations (black points). As higher frequency bins and bin 1 are of very low magnitude, only bins 2–20 were plotted to reduce the impact of noise on the plot. Shape metric variation before and after PVD and retinal detachment was no greater than test–retest variation in the absence of any change in ocular condition