The goal of medical image processing is to assist a human expert (e.g., a radiologist) in the task of identifying and interpreting all of the useful information available in an image. A medical image interpretation system constructs an internal model of the anatomy contained in the image so that it can be processed in an anatomical domain rather than simply as an array of gray levels. This allows the image interpretation system to incorporate anatomical (expert) knowledge into its analysis of the image and go beyond conventional image processing techniques. A medical image interpretation system brings to bear knowledge that is specific to the imaging system andâor the anatomy being imaged. This knowledge can be used to assist in solving the medical imaging problem at hand, whether it be identification of tissue for measurement (e.g., tumor volume to assess response to therapy) or tissue classification (e.g., determining whether the tumor is benign or malignant).
The use of domain knowledge via image interpretation techniques is important both in segmenting anatomical structures and analyzing them afterwards. Image interpretation techniques become necessary in many situations where conventional image processing and segmentation techniques are not sufficient. Conventional techniques may be adequate when the target object is relatively homogeneous and surrounded by a homogeneous, and sufficiently different, background. In this case, the object is easily segmented. This sometimes occurs in medical images, for example, an isolated lung nodule surrounded by well-aerated lung, imaged using computed tomography (CT). However, there are many situations where two (or more) separate tissues or organs overlap in an image and have the same range of gray values (representing similar physical characteristics under the specific imaging modality used). For example, in conventional radiography, where many tissues overlap due to the projectional nature of the imaging modality, a lung nodule may not be easily segmented if it overlaps with rib shadows. Similar problems can also arise in 3D image data, such as in segmenting the kidneys in CT because they are connected to other organs via vascular structures and may also be in contact with adjacent soft tissue structures (e.g., liver or spleen). In these cases, conventional segmentation methods such as the edge-, region-, or histogram-based methods (Chapter 2) are inadequate as they cannot distinguish between these adjacent tissues. Further processing is required to accurately identify each tissue or organ. Similar situations arise when the imaging system has poor spatial resolution, poor contrast resolution, when the image is noisy, or when more systematic errors (e.g., field inhomogeneities in magnetic resonance imaging) exist. Under such conditions, additional information is required to identify tissues and organs during image interpretation.
Online access to SPIE eBooks is limited to subscribing institutions.