Patch-based feature maps for pixel-level image segmentation

In this paper, we describe the use of phase-invariant complex wavelet filters, coupled to a training process involving a small, high-quality training dataset, to build an image segmentation system capable of performing in very low signal-to-noise, and under conditions of strong object-background contrast change. The three main components of our approach are: i) a patch-based feature description of local phase-invariant orientation fields; ii) a priori ground-truth data; iii) a machine learning method, such as Multilayer Perceptron (MLP) or kernel-based Support Vector Machine (SVM), to build an accurate classifier that is customised to the segmentation problem. A key feature of the approach is that it may be easily retrained and is, therefore, more adaptable to different imaging modalities. A representation of phase-invariant local image orientation using geometric algebra is first introduced; this is important to the patch-based approach. The quality of our trained systems is then assessed using Receiver Operating Characteristic (ROC) curves in two different biomedical applications: the human retinal vessel-bed in colour fundus images from the publicly available DRIVE database, and the rabbit endothelial cell boundaries of thoracic aorta microscopy images.