1 Background
2 Related work
3 Mathematical background
3.1 Preprocessing
3.1.1 Maximal overlap discrete wavelet transform
3.1.2 Denoising by MODWT
3.2 Contourlet transform
3.3 Hidden Markov random field model
-
Random field: The random variables are the intensity levels in an image. In HMRF model, two random fields exist:
-
Hidden random field: \(X = {(x_{1},x_{2},..,x_{N}) \Vert x_{i} \in L, i \in S}\) is a random field in a finite state space, L, and indexed by a set, S, with respect to a neighboring system of size, N. The state of this field X is unobservable/hidden and every \(x_{i}\) is independent of all other \(x_{j}\). The objective of this assumption is to classify each pixel independently (Marroquin et al. 2003).
-
Observable random field: \(Y = \{y = (y_{1},y_{2},..,y_{N}) \Vert y_{i} \in D, i \in S\}\) is a random field in a finite space, D, and indexed by a set, S, with respect to a neighboring system of size, N. This random field, Y, is observable and can only be defined with respect to X, where \(y_{i}\) follows a conditional probability distribution given any particular configuration \(x_{i} = l\): \(p(y_{i}\Vert l) = \{f(y_{i}; \theta _{l}), \ \forall l \in L \}\), where \(\theta _{l}\) is the set of the parameters that are involved.
-
-
Parameters: The set of parameters, \(\theta _{l}\), are unknown. Therefore, a model fitting solution is adopted to estimate them. In our context, the parameters are mainly the mean, \(\mu\), and the standard deviation, \(\sigma\).
-
Conditional independence: The two random fields, (X, Y), are conditionally independent.
-
Neighborhood system: It is a way to define the surrounding pixels for a specific pixel (Chen et al. 2011).
-
Clique: It is a subset of pixels, where every pair of distinct pixels are neighbors. A value is assigned to each clique, c, to define the clique potential \(V_{c}(x)\), where the sum of all of these values results in the energy function, U(x), that we aim to minimize.$$\begin{aligned} U(x) = \sum _{c \in C} V_{c}(x) \end{aligned}$$(4)
3.4 Modifications to conventional k-means clustering
4 Proposed segmentation algorithm
5 Datasets and results
5.1 Datasets
5.2 Results
5.2.1 Registration
5.2.2 Objective evaluation
Performance | Definition | Equation |
---|---|---|
Metric | ||
Accuracy | Correctness of the overall | \(\displaystyle \frac{ \text {TP} + \text {TN} }{ \text {TP} + \text {TN} + \text {FP} + \text {FN} }\) |
Segmentation | ||
DSI | Amount of overlap between | \(\displaystyle \frac{ 2 \times \text {TP} }{ 2 \times \text {TP} + \text {FP} + \text {FN} }\) |
The two segmentation | ||
FPR | Number of pixels incorrectly | \(\displaystyle \frac{ \text {FP} }{ \text {FP} + \text {TN} } = 1 - \text {Specificity}\) |
Segmented | ||
FNR | Number of pixels incorrectly | \(\displaystyle \frac{ \text {FN} }{ \text {FN} + \text {TP} } = 1 - \text {Sensitivity}\) |
rejected | ||
Sensitivity | Number of pixels segmented | \(\displaystyle \frac{ \text {TP} }{ \text {TP} + \text {FN} }\) |
Correctly | ||
Specificity | Number of pixels excluded | \(\displaystyle \frac{ \text {TN} }{ \text {TN} + \text {FP} }\) |
Correctly |
Measure | Dataset 1 | Dataset 2 | Dataset 3 | Dataset 4 | Dataset 5 | Dataset 6 | Average |
---|---|---|---|---|---|---|---|
Accuracy (%) | 99.99 | 99.92 | 99.78 | 99.75 | 99.28 | 99.15 | \({\text {99.72}}\) |
DSC (%) | 97.46 | 93.15 | 94.13 | 89.56 | 94.16 | 89.59 | \({\text {93.52}}\) |
FPR (%) | 0.01 | 0.06 | 0.17 | 0.08 | 0.16 | 0.04 | \({\text {0.07}}\) |
FNR (%) | 1.16 | 3.58 | 2.8 | 13.56 | 3.50 | 8.67 | \({\text {5.23}}\) |
Sensitivity (%) | 98.84 | 96.42 | 97.20 | 86.44 | 97.95 | 86.99 | \({\text {94.77}}\) |
Specificity (%) | 99.99 | 99.94 | 99.83 | . 99.92 | 99.77 | 99.84 | \({\text {99.96}}\) |
5.2.3 Subjective evaluation
Observation by | Dataset 1 | Dataset 2 | Dataset 3 | Dataset 4 | Dataset 5 | Dataset 6 | Average |
---|---|---|---|---|---|---|---|
Observer 1 | 5 | 4 | 4 | 3 | 4 | 4 | 4.00 |
Observer 2 | 5 | 4 | 5 | 3 | 5 | 4 | 4.41 |
Observer 3 | 5 | 4 | 5 | 4 | 5 | 3 | 4.16 |
Observer 4 | 5 | 4 | 4 | 4 | 4 | 4 | 4.08 |
Observer 5 | 5 | 4 | 5 | 3 | 5 | 3 | 4.08 |
Average | 4.14 |
6 Discussion
Region | Dataset 1 | Dataset 2 | Dataset 3 | Dataset 4 | Dataset 5 | Dataset 6 | Average |
---|---|---|---|---|---|---|---|
ROI (s) | 32.12 | 28.56 | 93.46 | 57.56 | 93.61 | 58.41 | 68.31 |
(51 slices) | (51 slices) | (161 slices) | (80 slices) | (155 slices) | (88 slices) | (110 slices) | |
Volume (min) | 4.38 | 3.21 | 3.30 | 5.31 | 3.90 | 5.87 | 4.60 |
(385 slices) | (385 slices) | (385 slices) | (385 slices) | (385 slices) | (385 slices) | (385 slices) |
Method | Accuracy | DSC | FPR | FNR | Sensitivity | Specificity |
---|---|---|---|---|---|---|
PCA-based Dakua et al. (2016) | 94.27 | 91.10 | 0.16 | 5.52 | 87.81 | 93.04 |
Global optimization Lawonn et al. (2019) | 94.98 | 92.08 | 0.15 | 5.45 | 90.42 | 94.62 |
Elastica Chen et al. (2020) | 95.07 | 92.91 | 0.12 | 5.33 | 92.20 | 97.24 |
Geometric Rai et al. (2021) | 95.11 | 92.93 | 0.12 | 5.31 | 93.10 | 97.38 |
Proposed method | 99.72 | 93.52 | 0.07 | 5.23 | 94.77 | 99.96 |
Method | Accuracy | DSC | FPR | FNR | Sensitivity | Specificity |
---|---|---|---|---|---|---|
SyN Avants et al. (2008) | 92.42 | 87.86 | 0.19 | 6.12 | 87.12 | 92.93 |
LT-Net Wang et al. (2020) | 94.57 | 88.98 | 0.16 | 5.76 | 89.23 | 94.12 |
VoxelMorph Balakrishnan et al. (2019) | 99.07 | 92.98 | 0.11 | 5.37 | 94.01 | 99.14 |
Dense-PSP-Unet Ansari et al. (2022) | 99.09 | 92.99 | 0.10 | 5.35 | 94.04 | 99.15 |
Proposed method | 99.72 | 93.52 | 0.07 | 5.23 | 94.77 | 99.96 |
Model name | Parameter count | Model size (MB) |
---|---|---|
SyN Avants et al. (2008) | 32,478,291 | 357.0 |
LT-Net Wang et al. (2020) | 29,478,291 | 331.2 |
VoxelMorph Balakrishnan et al. (2019) | 31,024,163 | 349.0 |
Dense-PSP-Unet Ansari et al. (2022) | 13,882,087 | 143.0 |
Proposed method | 13,289,122 | 143.0 |