Automatic Quantification of Changes in Corneal Neovascularization in Patients with Stevens-Johnson Syndrome

Abstract

Purpose : To develop and test an automated image processing pipeline for quantification of changes in corneal neovascularization (CNV) in patients with Stevens-Johnson syndrome (Figure 1), resulting from administration of bevacizumab and sunitinib over a 6 months period of treatment.

Methods : 1. Color images of ophthalmological images of corneas acquired using a slit-lamp microscope (TOPCON SL-D7 DC-3 8MP). Total set of 18 images.
2. Automated segmentation of vascularized pixels using the program B-Cosfire, version 1.4.0.0 (Azzopardi et al., 2014), software written in MatLab R14.
3. An ophthalmologist produced manually segmented gold standard reference images for each image.
4. Since B-Cosfire classifies all pixels along the border as vascularized (Figure 2b), a circular mask was determined which removes this border (Figure 2c). The automated method calculated the vascularized pixels as V_M and the reference vascularized pixels as V_R.
5. Dice similarity measure (DSM, Macedo et al., 2015) was used as a metric to evaluate the accuracy of the automated method compared to the manually segmented gold standard. DSM is computed by taking the intersection of V_M with V_R divided by the average of V_M and V_R.

Results : Mean DSM for the 18 images was found to be 0.369, with standard deviation of 0.048, and the maximum value of DSM was 0.423.
Comparing Figures 2a and 2c we notice that our method does not reproduce the very fine vessel structures in the reference segmentation, tending to exaggerate the thickness of the vessels. We believe that with better adjustment of the method parameters we will be able to reduce the width of the segmented vessels and thereby improve the measure.

Conclusions : The proposed image processing scheme was used to quantify the extent of CNV in color images of affected eyes. Our program, although the initial deviation, showed stable results (low std deviation) since it is a deterministic algorithm. We would expect it to allow assessment of individual patients with Stevens-Johnson syndrome through periodic examinations.

This is a 2020 ARVO Annual Meeting abstract.

 

Figure 1: (a) photo-image of the cornea; (b) manual segmentation of vascularized cornea; (c) extraction of vascularized pixels for vessel area calculation (Wakamatsu, 2014).

 

Figure 2: (a) Manually segmented gold standard image of corneal vascularization; (b) Automated segmentation using B-Cosfire; (c) Automated segmentation after removal of circular border pixels with mask.