Computer - aided grading of tumors using image analysis and supervised learning algorithms

Abstract

In the present dissertation our purpose was to improve the level of accuracy of thediagnostic and prognostic values associated with morphological malignancy grading of tissue biopsies of urine bladder cancer. Microscopic visual analysis of histopathological material provides an index of disease severity and tumour grading according to the degree of malignancy determines the choice and form(s) of treatment. Although histological grade has been shown to have prognostic significance in a wide variety of human neoplasms, the usefulness of tumour grading has been limited by poor inter and intra observer reproducibility. To encode and automate the way that experts perform pattern recognition tasks in their usual observation of tissue samples we designed a grade classification system employing histological/cellular features used by pathologists in assessing tumor grade. The system provided valuable information, indicating the importance of selected features in grade diagnosis and refining the subjective criteria for each category of tumour malignancy. To address the problem of grade characterization in a more objective and reproducible way we developed an image analysis system for automatic grade-classification, employing quantitative nuclear features. The usefulness of the system is its ability to automatically segment the nuclei on digitized tissue sections, following a routinely used staining procedure, and to provide diagnostic information with high accuracy. The accuracy of grade-classification system was further increased when we integrated all the available information in a so-called hybrid expert system, combining quantitative data with parameters derived by pathologists, indicating the complementary roles of an expert and an automatic image analysis system. Finally, to handle the difficult task of predicting cancer recurrence we implemented a prognostic system incorporating quantitative nuclear features. The system performed reasonably well, separating cases in to two prognostic groups: those who experienced recurrence and those who had no recurrence during the observation time. Concluding the results from the present dissertation are very promising in the field of computer-aided grade diagnosis and prognosis, raising the prospects for systems that would be able to associate input variables (images, clinical data) with clinically valuable output variables (diagnosis, prognosis).