Detection of Objects by Integrating Watersheds and Critical Point Analysis

This paper presents an improved method for detection of “significant” low-level objects in medical images. Information derived from watershed regions is used to select and refine saddle points in the discrete domain and to construct the watersheds & watercourses (ridges and valleys). The method overcomes previous topological problems where multiple redundant saddle points are detected in digital images. We also demonstrate an improved method of pruning the tessellation from which salient objects are defined. Preliminary evaluation was based on theoretical analysis, visual inspection of a set of medical images, and human observer experiments with promising result.