Liver-tumor boundary detection: human observer vs computer edge detection

As a preliminary step in computing tumor volume, we developed a computer edge detection program to define the liver-tumor interface in computed tomography (CT) images. Computer program performance was tested using CT images from a lucite liver/tumor phantom; from normal livers containing computer-generated pseudotumors of known size, object contrast, and liver-tumor edge gradients; and from 12 abdominal livers containing 19 focal tumors, eight with well-defined and 11 with ill-defined borders. Calculated sizes of the tumor phantom and pseudo-tumors were compared with measured volumes and predetermined cross-sectional areas, respectively. In the absence of a truth standard for the size of the focal hepatic tumors, computer-calculated cross-sectional areas of the tumors were compared with the measurements made by an experienced interpreter of CT images using the trackball cursor at the CT console. The console measurements were made five times on separate days during a one-week period. The variability in the measured areas of these tumors averaged 7.1% for the well-defined tumors and 14.0% for the poorly defined tumors (P = 0.05). The edge-linking algorithm systematically overestimated the volumes of individual slices of the hemispherical tumors in the lucite phantom. Nevertheless, because of algorithm failure in the slices containing the poles of the hemispheres, errors in total tumor volumes were -2.1% for the 5.1 cm radius tumor, +1.2% for the 2.7 cm radius tumor, and +15% for the 1.8 cm radius tumor. The edge-linking algorithm was reasonably successful in calculating areas of pseudotumors with object contrast of 3.0% or greater and steep edge gradients.(ABSTRACT TRUNCATED AT 250 WORDS)