Step 1: Load data

• we have 20 axial slices for the brain starting from the neck up to top of the head for each of three modalities: T1, T1C, T2

Step 2: 2-dimensional noise reduction (2D NR)

• Apply SUSAN (Smallest Univalue Segment Assimilating Nucleus) filtering independently on each slice

Step 3: Inter-Slice Intensity Variation Reduction (ISC)

• Use weighted regression to make tissue types consistently appear the same on different slices

Step 4: Intensity Inhomogeneity Reduction (INH)

• N3 (nonparametric intensity nonuniformity normalization) method enforces consistent tissue appearance within the same volume

Step 5: 3-dimensional noise reduction (3D NR)

• Apply SUSAN filtering to the entire 3d volume

Step 6: Coregistration (Coreg)

• Use normalized mutual information (NMI) to determine how best to align the different modalities

Step 7: Template Registration (Regist)

• Apply affine + non-linear transformations with different levels of regularization to align patient scans with “normal” template

Step 8: Intensity Standardization (Int Std)

• Use weighted regression to increase consistency within volumes of identical tissue identified by the template

Step 9: Feature Extraction + Pixel Classification

• Classify each pixel as being either “tumour” or “normal”

Step 10: Relaxation (Post Process)

• Correct potential mistakes made by the classifier (remove outliers that are unlikely to actually be tumour, smooth edges, fill holes, etc.)