MRI is becoming the primary imaging modality for liver oncological assessment and diffuse liver disease evaluation such as fibrosis and fatty liver disease [1]. Wide range in body habitus and also variations in liver shape and size between patients make it impossible to utilize a one size fits all protocol for patients. Technologists therefore end up customizing protocol for each patient and in the process lose valuable scanner time and patient comfort is compromised with longer exam times. This also leads to operator dependent results and could result in inconsistent image quality due to operator errors. This work focuses on addressing these challenges by automating the scan coverage and field of view (FOV) using the localizer/scout images.MRI Data: MRI data for this study were obtained on 1.5T GE Optima MR 450w and 3T GE DISCOVERY MR750 systems (GEHC) from 21 healthy volunteers. MR scout images was acquired using tri-planar 2D SSFSE sequence to generate 6-11 coronal, 5-10 axial and 2-6 sagittal slices. Parameters were: TE/TR = 80-83ms/435-1500 ms, FA = 90°, FOV = 410 to 480 mm2, TH = 8-10 mm, 512x512 matrix. Localization: Localizing the organ of interest involves feature generation followed by detection. a. Feature Generation: Histogram of Oriented Gradients (HOG) has in the past been used for object detection in 2D images [2]. As the name suggests, information contained within a patch is represented as a histogram of gradient orientations. In this instance, the HOG has been extended to 3D, which means that the gradient orientation at any voxel location has both an azimuth and an elevation component making it a 2D histogram. Additionally, we divided the patches into left-right and top-bottom cells and computed the HOG vector separately for each of the portions (see Figure 1). Since region descriptors based on histograms lose the spatial context, the division of patches into regions for feature vector computation and their subsequent concatenation provides a semblance of spatial arrangement for detection. b. Detection: We used a learning based algorithm to detect the bounding box of interest. Random forest classifier [3] was used to create a model of the liver bounding box from a set of training images. The model was used to detect the liver bounding box on a set of previously unseen images. The organ model obtained from the training provides a matching probability score for each of the evaluated patch locations on the test images. To test the performance of the algorithm, we computed the distance of the bounding box from the most superior and most inferior points of the liver surface. We used 16 randomly selected MR scout volumes for training and remaining 5 scout volumes were used for testing. The test images were chosen randomly and typically FOV ranged from top of the lungs to the pelvis. A trained radiologist marked the bounding box around the liver from which we computed the HOG feature vector in the training phase. Also, we randomly extracted a set of regions in the training images (excluding liver) to obtain negative examples. The feature vectors obtained were fed to the random forest classifier along with their labels and generated a model for testing. For detection of the liver bounding box in test images, we used the average bounding box size from training to sweep through the image and computed HOG feature for each patch location at different scales. With an appropriate threshold for the probability score we obtained the centroid of the liver bounding box. The execution time taken for detection was approximately 100 ms.In all the cases, we could always localize on liver FOV (Figure 2-4). On 4 test images with no anomalous anatomy, the localization error near the liver dome ranged between 0.14 c m and 3.4 cm, while the error near the liver apex ranged from -0.4 cm to 1 cm (see Figure 2). In Figure 3, subject had excess fat in and around the pericardium (white arrow), which created a curved surface with a lung above. This is exactly similar to what the algorithm has been trained for (i.e. mimic’s diaphragm) and as a result, error for bounding box on superior was large (4cm).We demonstrate that using fast scout images, we can achieve very good localization of liver FOV, irrespective of anatomy differences and hand-up / hands-down positioning.