| Image processing techniques were developed for pathological tissue analysis. The work included three main parts: analysis of tissue images stained with Perl's Prussian blue, development of object density-based segmentation algorithms for tissue image segmentation, and analysis of tissue images stained with hematoxylin and eosin.; For Perl's Prussian blue-stained tissue images, the blue areas were segmented based on color attributes. A series of image features were extracted to describe the subjective concept of “blueness”, an important attribute for pathological evaluation of blue-stained tissues. The features were selected through statistical analysis. Both statistical and neural network models were developed to predict expert pathological scores from the image features. The neural network model predicted pathological scores to an R2-value of 0.86.; Three object density-based algorithms were developed to segment images according to the spatial density of objects such as nuclei. The three algorithms were respectively based on primitive growing, wavelet transform, and influence zones. The algorithms were tested with both synthesized and real images. The test results showed the algorithms were effective, with the primitive-grow-based method performing better than the other two.; For hematoxylin and eosin-stained spleen sections, red pulps were segmented from white pulps based on differences in color and the lymphocyte density between red pulps and white pulps. Image features corresponding to a number of pathological attributes were extracted. Statistical and neural network methods were used to develop models for prediction of a comprehensive pathological score. The R2 value of prediction was 0.64 for a regression model and 0.75 for a neural network model. The neural network model showed advantage in comprehensive pathological score prediction. |
