Diffusion Maps Clustering for Magnetic Resonance Q-Ball Imaging Segmentation

Figure 2

N-Cuts generates overclustering due to sampling frequency variation in ODF images. In both eigenvalue plots Figures 2(a) and 2(d), the slope between the line joining 𝜆₀ and 𝜆₁ and the line joining 𝜆₁ and 𝜆₂ changes drastically, expressing an elbow in 𝜆₁, which indicates two clusters. The clustering results with Figures 2, 2(b), 3, 2(c), clusters are shown. Diffusion maps correctly finds two clusters, the object and the background, Figure 2(e). In the labeling, the ODFs are overlaid on the labels.

(a) N-Cuts eigenvalue plot

(b) N-Cuts, 2 clusters (blue and black)

(c) N-Cuts, 3 clusters (blue, orange, black)

(d) Diffusion maps eigenvalue plot

(e) Diffusion maps, 2 clusters (orange and black)