Digital image analysis has proved to be a powerful tool for providing a prognosis for cancer patients. For the prognosis to be as robust and reliable as possible, information regarding cell-type is needed, and damaged or overlapping nuclei have to be removed. Manually labeling the cell nuclei is time-consuming and expensive. An automatic labeling procedure would be an important contribution to the preprocessing of cell nuclei. In this thesis, we have developed a model for automatic classification of cell-type and removal of debris, using modern machine learning techniques. An investigation of the manual labeling of a set of experts is performed, to evaluate the performance of our approach. For removal of different types of debris we have developed highly specific novel features. We have also evaluated a set of previously known features, for use in cell-type classification. We generally found that automatic classification can achieve similar perfor- mance to that of human experts. The best results were found to be a correct classification rate of 97 % for cell-type classification and 87 % for the complete classification of both cell-type and debris. On the same small dataset used for evaluation of the human experts we found an average correct classification rate of 79.43 %. This result was better than the worst performing human expert and within the 0.95 confidence interval (85.14 ± 7.29%). Our approach shows promising results for automatic labeling of cell nucleus images, but may still be less robust than human experts. Further investigation of the human performance is needed to conclude on whether the whole labeling process can be fully automated and in order to chart out a direction for the further development of the automatic procedure.