A fully automatic multiscale 3-dimensional Hessian-based algorithm for vessel detection in breast DCE-MRI

OBJECTIVES The objectives of this study were to develop a fully automatic method for detecting blood vessels in dynamic contrast-enhanced magnetic resonance imaging of the breast on the basis of a multiscale 3-dimensional Hessian-based algorithm and to evaluate the improvement in reducing the number of vessel voxels incorrectly classified as parenchymal lesions by a computer-aided diagnosis (CAD) system. MATERIALS AND METHODS The algorithm has been conceived to work on images obtained with different sequences, different acquisition parameters, such as the use of fat-saturation, and different contrast agents. The analysis was performed on 28 dynamic contrast-enhanced magnetic resonance imaging examinations, with 39 malignant (28 principal and 11 satellite) and 8 benign lesions, acquired at 2 centers using 2 different 1.5-T magnetic resonance scanners, radiofrequency coils, and contrast agents (14 studies from group A and 14 studies from group B). The method consists of 2 main steps: (a) the detection of linear structures on 3-dimensional images, with a multiscale analysis based on the second-order image derivatives and (b) the exclusion of non-vessel enhancements based on their morphological properties through the evaluation of the covariance matrix eigenvalues. To evaluate the algorithm performances, the identified vessels were converted into a 2-dimensional vasculature skeleton and then compared with manual tracking performed by an expert radiologist. When assessing the outcome of the algorithm performances in identifying vascular structures, the following terms must be considered: the correct-detection rate refers to pixels identified by both the algorithm and the radiologist, the missed-detection rate refers to pixels detected only by the radiologist, and the incorrect-detection rate refers to pixels detected only by the algorithm. The Wilcoxon rank sum test was used to assess differences between the performances of the 2 subgroups of images obtained from the different scanners. RESULTS For the testing set, which is composed of 28 patients from 2 different clinical centers, the median correct-detection rate was 89.1%, the median missed-detection rate was 10.9%, and the median incorrect-detection rate was 27.1%. The difference between group A and group B was not significant (P > 0.25). The exclusion of vascular voxels from the lesion detection map of a CAD system leads to a reduction of 68.4% (30.0%) (mean [SD]) of the total number of false-positives because of vessels, without a significant difference between the 2 subgroups (P = 0.50). CONCLUSIONS The system showed promising results in detecting most vessels identified by an expert radiologist on both fat-saturated and non-fat-saturated images obtained from different scanners with variable temporal and spatial resolutions and types of contrast agent. Moreover, the algorithm may reduce the labeling of vascular voxels as parenchymal lesions by a CAD system for breast magnetic resonance imaging, improving the CAD specificity and, consequently, further stimulating the use of CAD systems in clinical workflow.