Edge and curvilinear structures detection on medical images via unsupervised, adaptive and deep learning approaches

Abstract

Computer-aided-diagnosis (CAD) systems are very useful to help doctors in detecting various human diseases. To build a CAD system, several computer vision algorithms are required, particularly to handle object detection and segmentation tasks automatically. To develop object detection algorithms, edges and curvilinear structures detection tasks are typically required. However, performing these tasks manually is tedious, time-consuming and prone to human errors. In this thesis, we design computer algorithms for edge and curvilinear structures detection, particularly for the application of optical disc boundary and retinal vessel detection from fundus images. The algorithms are developed based on the mathematical function that can closely represent the edge and curvilinear structures behaviours. The algorithms can detect edge and curvilinear structures under an unsupervised framework and they also allow an implementation with a deep learning architecture. This provides meaningful insight for robust edge and curvilinear structures detection algorithms developments on other image modalities.