Abstract:
In vivo confocal microscopy (IVCM) is enabling the examination of the corneal micro architecture. The sub basal corneal nerve plexus is particularly interesting because changes in its morphology promise to indicate the progression of conditions such as diabetic neuropathy, keratoconus, as well as damage after surgery. In this thesis, we investigate corneal nerve detection algorithms for the automatic segmentation of corneal nerves from IVCM images. We rst review the current major methods, and then assess the performance of these methods on a test image database of normals. We nd a maximum average performance of 80.8% for the sensitivity and 0.64 for the false positive rate. The presence of confounding structures such as epithelial cells decreases the e cacy of these algorithms. Therefore, we propose a method for the detection of epithelial cells - that treats epithelium as an image texture. The performance of the algorithm is tested, resulting in an average accuracy of 80.9% for epithelium detection. Finally we use epithelium detection in combination with a modi ed ridge based detection technique to segment the nerves. We nd that this technique produces results with high sensitivity (86%) and relatively low false positive rate (0.24), an improvement over the originally proposed approach (62.6% sensitivity and false positive rate of 0.49), as well as existing methods for nerve detection. This demonstrates that the segmentation of nerves in IVCM images can be approached to as a multi class problem where di erent structures (such as epithelial cells) are treated as classes, and ridges are good indicators of nerves.