Automatic Localization and Discrete Volume Measurements of Hippocampi from MRI Data Using a Convolutional Neural Network
VOLUME 8, 2020 This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/ 91725 A. Basher et al.: Automatic Localization and Discrete Volume Measurements of Hippocampi From MRI Data FIGURE 1 . This figure describes data generation process for the proposed method. Three-channel 2-D patches with their corresponding segmented labels are generated for the volume measurement. The left and right hippocampal voxels'
... pocampal voxels' locations from T1-weighted MRI scans were estimated using two-stage ensemble Hough-CNN. The estimated voxel location is used to extract 3-D patches with the size of 64 × 64 × 64. The extracted 3-D patches were then fed to the slicer to separate the slices for the axial, coronal and sagittal views. The axial, coronal and sagittal slices were normalized separately, rotated by -degree with a factor of n-times (in our case, n=10) and reshape into a size of 32 × 32 × 1. Similarly, from segmented label MRI scans, using the same voxel location and corresponding 3-D patches (size: 64 × 64 × 64) were generated. Slicer counts the number of pixels/voxels attributed in each slice for the axial, coronal and sagittal views. The 2-D slices from T1-weighted MRI scans were concatenated (along axis=2 (0-based axis)) to construct 3-channel 2-D patches with the size of 32 × 32 × 3. The 2-D patches with their corresponding labels (number of voxels/pixels) were fed to train the proposed CNN model. In the test phase, only the 2-D patches from the T1-weighted MRI scans are fed to the trained model to estimate the number of voxels/pixels attributed to each slice from the corresponding MRI scans. 91726 VOLUME 8, 2020