Figure 3
From: Deep learning for identifying corneal diseases from ocular surface slit-lamp photographs
The t-SNE visualization of the last hidden layer representations in the algorithm for diseases from the prospective dataset (510 images). Colored point clouds represent disease categories, showing how the algorithm clusters the diseases. Clusters of points represent our method’s ability to objectively separate normal patients from those with corneal diseases. Each point represents an ocular surface image projected from the 2048-dimensional output of the last hidden layer of the Inception-v3 backbone into two dimensions. We see clusters of points of the same clinical diseases. Patients with cataracts cluster in the center, while normal cornea cluster on the lower left. Corneal and limbal neoplasm cluster on the upper left. Among corneal disease, infectious keratitis is split across the corneal disease point cloud, indicating that it is prone to confusion with non-infectious keratitis and corneal dystrophy, which is in agreement with the confusion matrices results (Fig. 4).
