Computer-aided CT image analysis based on clustered hounsfield values

In Computed Tomography (CT) scanning, very small and controlled amounts of X-ray radiation are passed through different tissues in the body, which absorb radiation at different rates. After reconstruction of the attenuation coefficients using computers, CT numbers (called Hounsfield values) are displayed in a gray scale picture for interpretation by a radiologist, who is a specialized physician in CT and other radiology examinations. Manual interpretation, however, has its limitations as a high-speed CT scanner, such as four-dimensional CT, is commercially available. Thus, a computer-aided CT image analysis technique, specifically a physics-based technique, must reduce tae interpretation time and increase the accuracy. As a part of the larger project of building a "telematics-based customized cancer radiation treatment planning system," we developed semi-automatic watershed algorithms that classify pixels' Hounsfield values into regions by using mathematical morphology and digital topology. The process of clustering pixels in a medical image dataset labels them as anatomical structures with corresponding physiological properties. We applied our algorithms to a head phantom CT and to the CT data of a patient with hepatic metastases. We found that our segmentation tools were sufficient in providing anatomical structure definitions and radiological property calculations for the patient's CT data. Moreover, they were also useful in solving the challenging radiological problems such as accurately determining the extent and location of hepatic involvement and identifing a patient with metastatic liver disease. In addition, the hierarchical segmentation results were useful in extracting a region of interest without a priori anatomical knowledge of the human body. Compared with manual interpretation, this semi-automatic method decreased the processing time and increased the accuracy appreciably.