Automatic detection of abnormal vascular cross-sections based on density level detection and support vector machines.

PURPOSE: The goal is to automatically detect anomalous vascular cross-sections to attract the radiologist's attention to possible lesions and thus reduce the time spent to analyze the image volume. MATERIALS AND METHODS: We assume that both lesions and calcifications can be considered as local outliers compared to a normal cross-section. Our approach uses an intensity metric within a machine learning scheme to differentiate normal and abnormal cross-sections. It is formulated as a Density Level Detection problem and solved using a Support Vector Machine (DLD-SVM). The method has been evaluated on 42 synthetic phantoms and on 9 coronary CT data sets annotated by 2 experts. RESULTS: The specificity of the method was 97.57% on synthetic data, and 86.01% on real data, while its sensitivity was 82.19 and 81.23%, respectively. The agreement with the observers, measured by the kappa coefficient, was substantial (κ = 0.72). After the learning stage, which is performed off-line, the average processing time was within 10 s per artery. CONCLUSIONS: To our knowledge, this is the first attempt to use the DLD-SVM approach to detect vascular abnormalities. Good specificity, sensitivity and agreement with experts, as well as a short processing time, show that our method can facilitate medical diagnosis and reduce evaluation time by attracting the reader's attention to suspect regions.

Fast-marching contours for the segmentation of vessel lumen in CTA cross-sections.

This work deals with the segmentation of the arterial lumen in cross-sections of CT angiography (CTA) images, by means of active contours. Within the context of the fast-marching method, a new speed-control function is proposed in order to cope with strongly variable contrasts along the perimeter of the contour. This function was devised to guarantee the existence of a time T at which the fast-marching front fits the actual boundary of the vessel lumen, despite calcifications and other neighboring structures. Instead of using the magnitude of the image intensity gradient alone, this function includes exponential factors that strongly decrease the propagation speed when the front moves beyond the local maxima of the gradient magnitude and beyond the range of luminal intensities in CTA images. The propagation is stopped when the the growth of the area A encompassed by the front becomes very slow, which is characterized by a large value of dT/dA . The segmentation was evaluated in 65 cross-sections of carotid arteries from 13 different patients, by comparison with contours traced by a radiologist. The mean sensitivity was 0.849 and the mean positive predictive value was 0.797.