A morphologically driven gradient and marker controlled distance regularised level sets for nuclear segmentation in histopathological images

Abstract

The extraction of suitable biomarkers over a tissue image plays a vital role in the diagnosis and prognosis of cancer disease. Nuclear pleomorphism is one such trait, which serves as an important shape-based biomarker. An effective segmentation of the nuclei objects leads to an accurate diagnosis by an expert pathologist, which otherwise would be erroneous due to inter and intra-observer variability. In this research, a novel approach for segmenting the nuclei objects, using distance regularised level sets (DRLS), has been presented. It is shown that the shape prior based morphological transformation of the image achieves: a) centroid detection for accurate contour initialisation; b) gradient computation for an effective contour evolution. Experiments have been conducted on benign and malignant tissue images followed by a performance study using the object detection and the overlap resolution accuracy. Segmentation accuracy is assessed in comparison with the geodesic active contours, based on the ground truth.