1 Introduction
2 Vascular tree image synthesis
2.1 Vascularity model
-
number of output branches,
-
blood viscosity,
-
input and output blood flow,
-
input and output blood pressure,
-
perfusion volume.
-
matter preservation law,which couples flows (Q) of parent and child vessel segments of a single bifurcation (cf. Fig. 2),$$ Q_p = Q_l + Q_r, $$(1)
-
Poiseuille’s law,where \(\Updelta P\) denotes pressure drop on a vessel segment, η is the blood viscosity, R vessel radius, and l vessel length,$$ \Updelta P={\frac{8Q\eta l}{\pi R^4}}, $$(2)
-
bifurcation lawwhere γ is a bifurcation ratio—a factor that relates parent and child vessel radii.$$ R_p^{\gamma}=R_l^{\gamma}+R_r^{\gamma}, $$(3)
2.2 Imaging simulation
2.3 Specification of synthesized data sets
Vascularity parameter | Range | Number of classes | Total number of classes | Number of instances per class | |
---|---|---|---|---|---|
Per noise level | Per vascularity parameter | ||||
N
qinp
| 3,000–5,000 | 5 | 5 | 25 | 32 |
N
qout
| 3,000–5,000 | 5 | 5 | 25 | 32 |
η | 1–10 cP | 5 | 4 | 20 | 32 |
3 Texture analysis in 3D
-
Co-occurrence matrix (260 features),
-
Run-length matrix (25 features), and
-
Image gradient (5 features).
4 Dimensionality reduction
4.1 Feature space transformation
Vascularity parameter | Rice distribution parameter ν | ||||
---|---|---|---|---|---|
0 | 1 | 3 | 5 | 10 | |
Original feature space | |||||
N
qinp
| 36.9 (4.0) | 39.4 (3.5) | 54.3 (10.9) | 47.5 (8.3) | 53.8 (13.0) |
N
qout
| 24.4 (3.0) | 27.5 (3.0) | 28.1 (6.1) | 29.4 (4.1) | 33.1 (9.3) |
η | 0.0 (0.0) | 0.0 (0.0) | 0.8 (1.1) | 0.2 (0.3) | 0.3 (0.6) |
Principal component analysis | |||||
N
qinp
| 50.6 (4.5) | 51.9 (2.7) | 54.3 (3.5) | 57.4 (4.8) | 58.9 (5.6) |
N
qout
| 23.8 (9.5) | 29.4 (6.7) | 28.8 (10.1) | 28.8 (12.6) | 29.4 (13.0) |
η | 0.0 (0.0) | 0.0 (0.0) | 0.8 (1.1) | 0.2 (0.4) | 0.3 (0.6) |
Random projection | |||||
N
qinp
| 31.3 (6.9) | 38.1 (6.0) | 40.0 (7.2) | 49.4 (4.3) | 51.3 (6.5) |
N
qout
| 21.8 (9.1) | 24.4 (3.6) | 25.9 (11.5) | 20.1 (9.8) | 30.6 (14.1) |
η | 6.5 (1.4) | 7.4 (1.8) | 9.7 (3.1) | 2.9 (1.1) | 8.4 (2.8) |
4.2 Feature selection
4.2.1 Clustering stability
-
the same cluster both in \(\pi_1(S_1)\) and \(\pi_2(S_2),\)
-
the same cluster in \(\pi_1(S_1)\) but to different clusters in \(\pi_2(S_2),\)
-
different clusters in \(\pi_1(S_1)\) but in the same cluster in \(\pi_2(S_2).\)
4.2.2 Clustering stability-based feature selection
5 Experiments
5.1 The experimental framework
5.2 Discussion of the results
Evaluation criterion | Rice distribution parameter σ | ||||
---|---|---|---|---|---|
0 | 1 | 3 | 5 | 10 | |
Constant input flow | |||||
\(\varphi\)
| 0.0 | 26.3 | 34.4 | 46.3 | 53.1 |
\(\varphi(1-d_{\beta}^J)\)
| 2.5 | 3.1 | 9.4 | 39.4 | 40.0 |
\(\varphi(1-d_{\beta}^H)\)
| 0.0 | 3.1 | 6.9 | 31.3 | 45.0 |
\(\varphi\left({\frac{1}{1+d_\beta^M}}\right)\)
| 0.0 | 1.9 | 8.8 | 34.3 | 47.5 |
Constant output flow | |||||
\(\varphi\)
| 0.0 | 0.6 | 10.0 | 18.5 | 26.3 |
\(\varphi(1-d_{\beta}^J)\)
| 0.0 | 0.0 | 6.9 | 5.6 | 26.3 |
\(\varphi(1-d_{\beta}^H)\)
| 0.0 | 0.0 | 0.6 | 4.4 | 26.3 |
\(\varphi\left({\frac{1}{1+d_\beta^M}}\right)\)
| 0.0 | 1.8 | 5.6 | 6.9 | 28.1 |
Viscosity | |||||
\(\varphi\)
| 0.0 | 0.0 | 0.8 | 0.0 | 0.0 |
\(\varphi(1-d_{\beta}^J)\)
| 0.0 | 0.0 | 0.8 | 0.0 | 0.0 |
\(\varphi(1-d_{\beta}^H)\)
| 0.0 | 0.0 | 0.8 | 0.0 | 0.0 |
\(\varphi\left({\frac{1}{1+d_\beta^M}}\right)\)
| 0.0 | 0.0 | 0.8 | 0.0 | 0.0 |