Analysis of Hand-Crafted and Automatic-Learned Features for Glaucoma Detection Through Raw Circmpapillary OCT Images
Taking into account that glaucoma is the leading cause of blindness worldwide, we propose in this paper three different learning methodologies for glaucoma detection in order to elucidate that traditional machine-learning techniques could outperform deep-learning algorithms, especially when the image data set is small. The experiments were performed on a private database composed of 194 glaucomatous and 198 normal B-scans diagnosed by expert ophthalmologists. As a novelty, we only considered raw circumpapillary OCT images to build the predictive models, without using other expensive tests such as visual field and intraocular pressure measures. The results ratify that the proposed hand-driven learning model, based on novel descriptors, outperforms the automatic learning. Additionally, the hybrid approach consisting of a combination of both strategies reports the best performance, with an area under the ROC curve of 0.85 and an accuracy of 0.82 during the prediction stage.