We have used LONI ICE to locate the left and right eyes, nose, ventricles, central sulcus, and anterior and posterior commissures in three-dimensional MRI brain image volumes. After locating the ventricles, we also use LONI ICE to determine the MRI weighting of the image volume. The training and testing data sets for each anatomical landmark are listed in .
Image volumes in the training and testing sets
The ICE network shown in is trained to locate the subject’s right eye, and the ICE network shown in is trained to determine the MRI weighting (PD, T1, T2) of an image volume. We also trained two other ICE networks similar to that locate the subject’s left eye and the bridge of the subject’s nose. Each gray box in represents a trained ICE module and each white box displays the value of a three-dimensional coordinate (X
, or Z
) when the ICE network is run. The input point (in both ICE networks it is the center of a 256 × 256 × 256 cube) at the top of each ICE network is processed downwards through the ICE modules and an output point is returned at the bottom. The first two rows of text on each ICE module display the grid size and grid cell size used to average voxel intensities. ICE modules that use singular value decomposition to solve Eq. 6
display the calculated precision with a third row of text, and those without a third row of text use neural networks to separate image volumes into different classes.
Two networks of LONI ICE modules. a An ICE network trained to locate the subject’s right eye. b An ICE network trained to identify the weighting (PD, T1, or T2) of an MRI image volume
We used 64 T1-weighted and 64 T2-weighted image volumes to train the eyes and nose ICE networks and 16 T1-weighted and 16 T2-weighted image volumes for testing. The results for a T1-weighted (top) and a T2-weighted (bottom) image volume from the testing set are illustrated in . The solid white squares graphically depict how precisely the bridge of the nose and the right and left eyes are determined. Each eye is detected within a rectangular volume of (10, 10, 10) voxels and each nose bridge is located within a rectangular volume of (10, 12, 16) voxels. Even though the image volumes have different resolutions and the subject’s nose is cut off in the T1-weighted image volume, the anatomical point landmarks are still well-determined. In order to verify the accuracy of the right eye ICE network, we computed a histogram of the errors determined from 144 T1-weighted and 92 T2-weighted image volumes. The errors are defined as the differences in each dimension between the manually selected points and the points output by the ICE network. As illustrates, the output points are correctly localized to the error bounds [(−5, 5) in each dimension] computed by our approach.
Fig. 11 Automatic detection of the eyes and bridge of the nose in a T1-weighted (top) and a T2-weighted (bottom) image volume. The axial and sagittal view planes intersect the bridge of the subject’s nose. The solid white squares show how precisely the (more ...)
Error histograms for the right eye ICE network. The errors in all three dimensions are localized within the calculated error bounds of (−5, 5) voxels
As illustrates, our strategy for determining MRI weighting starts with two ICE modules that locate the center of the subject’s ventricles, and are followed by an ICE module that classifies the image volume as either PD-/T1-weighted or T2-weighted. When the image volume is classified as PD-/T1-weighted, an additional ICE module is used to further classify it as PD-weighted or T1-weighted. We use the subject’s ventricles for classification because the intensities of the voxels in that region are different for the three MRI weighting types. The MRI weighting ICE network was trained using 76 PD-weighted, 62 T1-weighted, and 72 T2-weighted image volumes and was tested using 20 PD-weighted, 20 T1-weighted, and 20 T2-weighted image volumes. shows the results for a PD-weighted (top), T1-weighted (middle), and T2-weighted (bottom) image volume from the testing set. The solid white squares show the center of each subject’s ventricles as output by the second ICE module in . All the image volumes from the training and testing sets were correctly classified.
Fig. 13 Automatic detection of the ventricles in a PD-weighted (top), T1-weighted (middle), and T2-weighted (bottom) image volume. Both the axial and sagittal views for each image volume are shown. The solid white squares show how precisely the center of the (more ...)
summarizes how well the trained LONI ICE networks can locate anatomical landmarks in MRI image volumes of adult heads. Included in this table are results for the anterior and posterior commissures. Only T1-weighted images were used for training and testing because of the difficulty in manually finding the commissures in T2-weighted images. There are also results for the lateral termination of the central sulcus (defined as the extrapolated intersection of the central sulcus with the Sylvian fissure on the surface of the right hemisphere). The central sulcus point was determined by manually removing the skull in each image volume and then generating a brain surface. Each point was chosen by viewing and manipulating each surface in a three-dimensional display.
Precisions of trained LONI ICE networks that locate anatomical landmarks in MRI images of adult heads