Optic Disc and Fovea Localization based on Anatomical Constraints and Heatmaps Regression

In this paper, we deal with anatomical landmark localization as a heatmap regression problem. Based on this, we introduce a lightweight architecture to simultaneously localize fovea and optic disc (OD). Additionally, considering that directly attaching argmax to the output layer can lead to confidence map offsets errors, we propose a centroid clustering algorithm to address this issue. Extensive experiments are constructed on the IDRiD dataset, confirming the superiority of the proposed method. In particular, the Euclidean errors on fovea and OD are 45.034 and 21.101 (in pixels), respectively, which exceeds the other competitors of IDRiD Challenge 2018 by a large margin. Furthermore, at a resolution of
$420\times 356$
the 90ms inference speed of a single image is conducive to large-scale clinical diagnosis.