2D CNN-Based Slices-to-Volume Superresolution Reconstruction

In the case of slice-to-volume reconstruction, the usual method is to head up the slices according to their relative positions. However, it is common that the space between two scanned slices is more than the interval of the pixels on each slice, leading to the result that the reconstructed volume is anisotropic. To obtain isotropic reconstruction, traditional methods attempt to interpolate missing slices based on interpolation algorithms. However, the resolutions of the reconstructed volume are still different in different directions since the slices are highly related. Moreover, this may lead to misplacement among the slices in the nonscanning directions of the volume. As such, we intend to overcome the above two obstacles on the basis of the 2D CNN superresolution strategy. In the proposed method, the superresolution is conducted for the nonscanned slices. Moreover, the samples for training the CNN are specifically designed, which have tomographic features similar to those of the nonscanned slices. Additionally, a multichannel and dense connected CNN is borrowed to perform slice superresolution considering that each slice is actually related to its neighbors. Finally, we conduct slice-to-volume reconstruction after performing superresolution for slices. Experiments in the case of single-channel and multichannel CNNs are conducted to perform validation. The proposed method is robust and can obtain high accuracy.