Comprehensive Retinal Image Analysis for Aggressive Posterior Retinopathy of Prematurity

Abstract

Computer aided analysis plays a nontrivial role in assisting the diagnosis of various eye pathologies. In this paper, we propose a framework to help diagnose the presence of Aggressive Posterior Retinopathy Of Prematurity (APROP), a pathology that is characterised by rapid onset and increased tortuosity of blood vessels close to the optic disc (OD). We quantify vessel characteristics that are of clinical relevance to APROP such as tortuosity and the extent of branching i.e., vessel segment count in the defined diagnostic region. We have adapted three vessel segmentation techniques: matched filter response, scale space theory and morphology with local entropy based thresholding. The proposed feature set equips us to build a linear discriminant classifier to discriminate APROP images from clinically healthy images. We have studied 36 images from 21 APROP subjects against a control group of 15 clinically healthy age matched infants. All subjects are age matched ranging from 33-40 weeks of post menstrual age. Experimental results show that we attain 100% recall and 95.45% precision, when the vessel network obtained from morphology is used for feature extraction.

Fig 1. Representative NN data set.

(a) Typical clinically healthy infant eye with immature vessels; (b) infant eye with apparent tortuous vessels in the nasal zone (left of the OD); (c) infant eye with developed vessels reaching the periphery of the eye; (d) severe APROP infant eye; (e) infant eye with APROP onset.

Fig 2. Proposed algorithm for vessel segmentation.

Fig 3. Stages of preprocessing.

(a) Original image with arrows indicating saturation noise; (b) binary mask to remove saturation noise; (c) outer ring mask overlaid & saturation noise removed; (d) markers indicating major axis end points used to compute OD mask; (e) circular OD mask.

Fig 4. Relative uniformity in illumination across the red, green and blue channels respectively.

(a-c) RGB color bands of a healthy fundus image, with blue being most uniform; (d-f) RGB color bands of an APROP fundus image, with the mean of blue channel <64.

Fig 5. Stages in image blurring and subtraction.

(a) vessel blurring; (b) subtraction from selected channel; (c) final segmented network; (d) segmented network overlaid in original image.

Fig 6. Vessel post processing.

(a) branch points and end points indicated in blue and red respectively; (b) overlaying branch points results in pruned vessels; (c) terminal spurs indicated in red; (d) ‘L’ and ‘T’ shaped critical points leading to false end points; (e) subsequent thinning and branch points recomputed; (f) portion of the vessel skeleton in gray box zoomed in through (a-e).

Fig 7. Pictorial representation of diagnostic regions (DR1, EDR1 & DR2).

Fig 8. Vessel network extracted from.

(a) matched filter response and (b) scale space segmentation.

Fig 9. Vessel segments before area threshold.

(a-c) prior to applying the area threshold (a) matched filter response; (b) scale space segmentation; (c) morphology with local entropy; (d-e) after the application of area threshold: (d) matched filter response; (e) scale space segmentation.

Fig 10. Anova plots.

(a-c) Anova plots comparing T feature in (a) matched filter response; (b) scale space segmentation; (c) proposed method and (d-e) Anova plots comparing S feature in (d) matched filter response; (e) scale space segmentation; (f) proposed method.

Fig 11. Result of LDA classification on OD centered images.

(a) OD DR1; (b) OD EDR1.

Fig 12. Vessel networks of a FN.

(a) Original APROP image; (b) vessel network from matched filter response; (c) vessel network from scale space theory; (d) vessel network from proposed morphology.

Fig 13. Vessel networks of a FP.

(a) Original healthy image; (b) vessel network from matched filter response; (c) vessel network from scale space theory and (d) vessel network from proposed morphology.

Fig 14. Healthy image overlaid with vessel segments in DR1 (white) and DR2 (black).

Fig 15. Varying levels of arborocity in APROP.

(a-b,d-e) vessel segments in DR1, EDR1 and DR2 respectively; (c) APROP sample without vessel segments in DR2; (f) APROP sample with too many vessel segments in DR2.

Fig 16. Varying levels of arborocity in healthy.

(a-b,d-e) vessel segments in DR1, EDR1 and DR2 respectively; (c) Immature clinically healthy sample; (f) Mature clinically healthy sample.

Fig 17. Severe APROP subject treated post diagnosis.

(a) vessels after first sitting of laser burns; (b) vessels with reduced tortuosity after second week of laser therapy.