In bolus tracking techniques monitoring the arrival of the bolus, the operator prescribes the 2D plane of MRA crossing the aorta and manually starts the dynamic contrast-enhanced (DCE) scan. The prescription of the 2D plane is a complex process, and its accuracy is highly dependent on the skill of the operator. Our goal is to position the 2D plane automatically, thus resulting in the improvement of operator workflow.To prescribe a 2D plane cross-section of the aorta, accurate detection of the aorta location is necessary. Ordinary 2D SSFSE scout images are analyzed to detect the aorta. Fig. 1 shows the flow of our algorithm. Firstly, axial images (> 6 slices) are classified by anatomy, including the lung, the boundary of the lung, the liver, and the liver with two kidneys. We used Laplacian eigenmaps¹ for this classification. The axial image is cropped, inscribed outside of the thorax, and then resized to a rectangular shape (44 × 22 pixels). Following classification, one of four atlases that correspond to the four classifications is applied to the axial image, which can narrow down the existing area of the aorta. The axial images are also used for the detection of cerebral spinal fluid (CSF) that is searched by the AdaBoost classifier and becomes an anchor point in searching for the aorta². To detect the aorta, we used the Hough Forests classifier³ that utilizes the feature value in the neighboring patches of objects, and creates a map of votes from each feature patch. The patch (16 × 16 pixels) is extracted from a sub-window (48 × 48 pixels) in the axial image, and the sub-window is rotated around the CSF to detect the aorta in the next axial image (Fig. 2). The Hough map is weighted by the atlas. The location of the maximum value in the map shows the aorta location. There are two aorta classes in SSFSE axial images. One appears bright and the other appears dark depending on whether flow spins are re-phased or de-phased due to the blood velocity. The Hough Forests classifier detects both the bright and dark aortas, and the AdaBoost classifier determines whether it is dark or bright². The esophagus, which is near the aorta, has a similar contrast and shape to the aorta. This can cause misdetections and false positives. We built a model expressing the location distribution of the aorta in relation to the esophagus (Fig. 3). With this model, another peak apart from the incorrectly detected peak in the Hough map is searched, and another peak is identified as the aorta peak if the peak value is more than 50% of the esophagus. The oblique plane is positioned by the line crossing the center of the detected aortas (7 aortas is needed at least) using the regression method (Fig. 1d). The oblique plane was reformatted from 3D datasets of DCE liver images acquired in the aortic phase to simulate an oblique MRA plane. We tested 40 patients (mean age = 64 years; age range = 24 – 85 years) after informed consent was obtained under institutional review and approval.The Laplacian eigenmaps consisted of 218 cropped images. The accuracy rate was approximately 80%. For aorta detection, 10,000 patches extracted from 200 axial images of 43 patient datasets were used as the training step for the Hough Forests classifier. Aside from these learning datasets, 358 axial images were used for testing from 40 patients. Table 1 shows the error of the detected center position of the aorta, which was calculated and compared to the manual results (ground truth). In the resulting reformatted oblique images, 5 of 40 datasets did not show 50% of the cross-section of the aorta. Fig. 4 shows an example of the 2D oblique planes for success and failure cases.In aorta detection, most of the error was caused in the detection of the bright aorta because of the difficult identification of the center of the aorta. In particular, the occurrence frequency of the bright aorta in patients tended to be greater than in volunteers that we tested previously² due to slow vessel flow. Therefore, it is critical to decrease the occurrence of the bright aorta.We proposed a new method for automated positioning of MRA for bolus tracking, and demonstrated that our algorithm was able to plan the 2D scan plane for MRA. This automation will help the operator and decrease the total study time.