| dc.description.abstract | In Medical Imaging of the brain, especially Magnetic Resonance Imaging (MRI), localizing various anatomical landmarks like the Anterior and Posterior Commissure (A.C./P.C.) and Mid-Sagittal plane (MSP) is crucial for good quality MRI. By convention, during brain MRI scan acquisition, the radiographer first performs a three-plane MRI localizer slice acquisition protocol to obtain these landmarks. This process is called Scout Scan. However, this is a tedious job and is susceptible to operator error. Also, the MRI scan�s resolution is anisotropic, i.e., good in-plane and lower out-of-plane resolution. As a result, a change in head position might significantly impact the interpretation of an MRI \ image. Hence, Automizing this process is vital to reduce operator error.Previous works predict A.C. and P.C. points in the Mid-sagittal plane, but the improper head position may lead to an improper Mid-sagittal plane (MSP). Hence, it may lead to localization errors. Also, previous works predict this point in a 3D voxel, which is impractical. To obtain a 3D voxel, longer time and computational resources are required. Furthermore, G.E.�s healthcare system has developed a similar tool named �AIRx, but it takes 9 to 20 Localizer slices to predict these landmarks.This work presents the deep learning-based automated localizing of these landmarks in 3D space for the brain from a three-plane 2D MRI localizer slice. This work uses six publicly available brain MRI datasets and a few image augmentation techniques. The mean error in localization of A.C. and P.C. within the dataset is less than 1mm. For cross dataset, it is less than 2mm, and also mean error in degrees for finding orientation vector is less than 2� for both within and cross dataset. | |
