Analysis of soft tissue tumors by an attributed minimum spanning tree.

Histologic slides of 22 soft tissue tumors (9 malignant fibrous histiocytoma, 8 fibrosarcoma, 2 rhabdomyosarcoma, 2 osteosarcoma, 1 Askin tumor) were Feulgen stained. Using an automated image analyzing system (Cambridge 570) at low magnification (25x), the tumor cell nuclei were segmented. The geometrical center of the nuclei was considered the vertex. A basic graph was constructed according to the neighborhood condition of O'Callaghan. Neighboring tumor cell nuclei were visualized by connecting edges. Several features of tumor cell nuclei were measured, including area, surface, major and minor axis of best fitting ellipsis and extinction (DNA content). Nuclear features are attributed to the vertices. The differences, or "distances," between features of connected vertices are attributed to the corresponding edges, which are dependent on the attributes. Thus, different minimum spanning trees (MST) result. Each MST can be decomposed into clusters using a suitable decomposition function on the edges, which rejects an edge if its attributes differ from the mean of the attributed values of surrounding edges more than a neighbor dependent bound (lower limit). Taking into account the length and other attributes of edges (e.g., differences in orientation of the major axis), clusters of different nuclear orientation can be detected. A cluster tree can be constructed by defining the geometric center of a cluster as a new vertex, and by computing the neighborhood of the cluster vertices. The result is an attributed MST containing characteristic structural properties of the image (in cases of sarcomatous tumors, local orientation of tumor cell nuclei and local DNA abnormalities).