Superpixel Based Automatic Segmentation of Corneal Ulcers from Ocular Staining Images

In this paper, we proposed and validated a novel and accurate automatic pipeline for extracting flaky corneal ulcer areas based on fluorescein staining images. We first used an existing semi-automatic approach to identify the cornea from each image. The ulcer area was then segmented within the cornea by employing a combination of techniques: 1) identify and modify the color information of reflective areas; 2) segment each image into a total of 1000 superpixels based on simple linear iterative clustering (SLIC); 3) employ support vector machine (SVM) to classify all superpixels into two classes; 4) erode and dilate to polish the ulcer segmentation results. The proposed pipeline has been validated on a total of 150 clinical images. Accurate segmentation results have been obtained, with the mean accuracy being 0.984, the mean Jaccard similarity coefficient being 0.871, and the Pearson correlation coefficient being 0.921 when compared with the manually-delineated gold standard. The proposed method was found to significantly outperform two classic segmentation algorithms (active contour and Otsu thresholding) in terms of segmenting corneal ulcers.