Deep Adaptive Log-Demons: Diffeomorphic Image Registration with Very Large Deformations

Algorithm 3

Two-layer unsupervised deep adaptive registration framework.

Classifier Training: Eight-layer CNNs trains different levels of rotation and deformation images

Registration:

Input: Images , and initial velocity field

Output: Transformation from to

Repeat

Pre-registration: For 2D image, through CNNs classifier to rectify rotation then scale and transformation;

For 3D image, triplanar 2D CNNs to pre-registrate

Demons registration:

 (i) Find updates mapping to

 (ii) Smooth updates:

 (iii) Update velocity field:

      (approximated with )

 (iv) Smooth velocity field:

 (v) Use PCA to extract features and calculate new similarity metric as PCA-SSD, PCA-pearson,

    PCA-spearman, and kendall.

until convergence