A threshold method based on statistical principles and heuristics is developed to detect boundaries in radiographic images. Each local region of the image containing a portion of boundary is characterized by a mixture of two (normal) intensity distributions. Thresholds are set dynamically according to local, rather than global, characteristics estimated from the observed intensity histograms. A program to implement the method has been written. Experimental results on cardioangiograms are presented to successfully demonstrate the feasibility of the method for low quality images. The method is insensitive to shading or gradually varying interference. © 1972.