Segmentation of chest radiographs as a pattern classification problem
Michael F. McNitt‐Gray · 1993
In digital chest radiography, the goal of segmentation is to automatically and reliably identify anatomic regions such as the heart and lungs. Aids to diagnosis such as automated anatomic measurements, methods that enhance display of specific regions, and methods that search for disease processes, all depend on a reliable segmentation method. The goal of this research is to develop a segmentation method based on a pattern classification approach. The pattern classification approach consists of classifying each pixel into one of several anatomic classes on the basis of one or more feature values. In this research, three types of locally calculated features are used: gray-level based measures, local difference measures and local texture measures. Three classifiers are used: a linear discriminant function, a k-nearest neighbor approach and a neural network. Supervised techniques train each classifier to learn the characteristics of the anatomic classes. Each classifier is trained and tested using normal chest radiographs. Using a test set of 16 chest images, the linear discriminant correctly classified 70.6% the k-nearest neighbor correctly classified 71.0% and the neural network correctly classified 76.9% of the pixels. An additional step was added to account for spatial information. The combined steps achieved 87.0%, 87.1% and 88.8% correct for the three classifiers, respectively. This approach was also tested on CT image slices. Using locally calculated features alone achieved 84.7%, 87.4% and 87.2% correct for the three classifiers over all CT test images. The combined steps were applied to one CT image slice and the method achieved 95.3%, 96.2% and 96.3% correct for the three classifiers. The pattern classification approach to image segmentation has shown promise for further development. Locally calculated features are important in classifying pixels, but these alone may not be sufficient. A method for incorporating spatial information into the classification decision appears to improve the results and may be necessary for reliable segmentation. This dissertation also shows that the pattern classification approach may be applied to images from other modalities.