Combining efficient hand-crafted features with learned filters for fast and accurate corneal nerve fibre centreline detection.

We propose a new approach to corneal nerve fibre centreline detection for in vivo confocal microscopy images. Relying on a combination of efficient hand-crafted features and learned filters, our method offers an excellent compromise between accuracy and running time. Unlike previous solutions using sparse coding to learn small filter banks, we employ K-means to efficiently learn the high amount of filters needed to cope with the multiple challenges involved, e.g., low contrast and resolution, non-uniform illumination, tortuosity and confounding non-target structures. The use of K-means for dictionary learning allows us to learn banks of 100 filters in less than 30 seconds compared to several days needed when using sparse coding. Experimental results using a dataset including 100 images show that our approach outperforms significantly state-of-the-art methods in terms of precision-recall curves.