| Summary: | We have developed a multistage image analysis technique for the simultaneous segmentation of blood vessels and hypoxic regions in dual-stained tumor tissue sections. The algorithm, which is integrated in a task-oriented image analysis system developed on-site, initially uses the K-nearest neighbor classification rule in order to label the image pixels. Classification is based on a training set selected from manually drawn regions corresponding to the areas of interest. If the output image contains a significant number of misclassified pixels, the user has the option to apply a series of specific problem-designed routines (texture analysis, fuzzy c-means clustering, and edge detection) in order to improve the final segmentation result. Validation experiments indicate that the algorithm can robustly detect these biological features, even in tissue sections with a very low quality of staining. This approach has also been combined with other image analysis based procedures in order to objectively obtain quantitative measurements of potential clinical interest.
|
|---|