2D CNN vs 3D CNN

“a detailed comparison between 2D and 3D neural networks for medical image recognition and show that 3D convolution neural networks (CNNs) are more effective and less likely to miss regions of interest in medical images.”

"because 2D CNNs take a single slice as input, they inherently fail to leverage context from adjacent slices. Voxel information from adjacent slices may be useful for the prediction of segmentation maps."

" The ability to leverage interslice context can lead to improved performance but comes with a computational cost as a result of the increased number of parameters used by these CNNs."

A fully-convolutional 2D CNN [2] was trained for segmentation of the scans. Slices were kept in the original resolution (256x256 pixels) and the model was trained using a batch size of 8. The model was trained using full 2D slices.

A 3D CNN [3] was trained for the same task. Settings used for the training of the 2D CNN were kept the same for the 3D CNN except for input size and batch size. The 3D CNN was trained on 3D patches from the training set with dimensions 256x256x8 and batch size of 1.

The main idea between 2D and 3D is 2D CNN uses only width and height as its input size and the batch_size as its total counts for each loop whereas 3D CNN uses (width, height, batch_size(originally were used for 2D CNN, now it becomes the third axial of each input) and the batch_size is now 1)

2D convs

3D convs