Automatic Corneal Ulcer Segmentation Combining Gaussian Mixture Modeling And Otsu Method

In this paper, we proposed and validated a novel and accurate pipeline for automatically segmenting flaky corneal ulcer areas from fluorescein staining images. The ulcer area was segmented within the cornea by employing a joint method of Otsu and Gaussian Mixture Modeling (GMM). In the GMM based segmentation, the total number of Gaussians was determined intelligently using an information theory based algorithm. And the fluorescein staining images were processed in the HSV color model rather than the original RGB color model, aiming to improve the segmentation results' robustness and accuracy. In the Otsu based segmentation, the images were processed in the grayscale space with Gamma correction being conducted before the Otsu binarization. Afterwards, morphological operations and median filtering were employed to further improve the Otsu segmentation result. The GMM and Otsu segmentation results were then intersected, for which post-processing was conducted by identifying and filling holes through a fast algorithm using priority queues of pixels. The proposed pipeline has been validated on a total of 150 clinical images. Accurate ulcer segmentation results have been obtained, with the mean Dice Similarity Coefficient (DSC) being 0.88 when comparing the automatic segmentation result with the manually-delineated gold standard. For images in the RGB color space, the mean DSC was 0.83, being much lower than that of the images in the HSV color space.