Table 1.
Participant characteristics and scanner type.
Fig 1.
BIC and AIC score displayed for the 10th (top) and 30th (bottom) bin in 50 bin.
Results are obtained by using pSMC normalization of the ADNI dataset. The AIC (red) and BIC (red) values are simply -2log(likelihood) (blue) plus 2P (green) and log(N)P (green), respectively. Refer to Eq 7 and Eq 8 for the definition of AIC and BIC. (a), (b) (d) and (f) show the AIC and BIC along with their components' value. (c) and (f) offers a direct comparison between AIC and BIC. The 10th bin contains 586 voxels, hence it is a small sample, in contrast to the large sample of the 30th bin containing 12,414 voxels. Note that the term “sample” in this context does not mean the number of patients, but the number of voxels to be clustered.
Table 2.
Grade of evidence of the BIC difference [35].
Fig 2.
Generalization error (GE) by BIC and AIC using pSMC normalization of MCI versus AD on the ADNI dataset.
The generalization error (equivalently, 1-accuracy) is computed based on features (mean and standard deviation value) extracted from the clusters in the individual bin (one bin among 50 bins, for example) via 10 times 10-fold cross-validation. For example, a green point can represent the classification GE using n-th bin when we divide the whole brain voxels into 50 bins. The plotted points are the bins that yield GE < 0.25, and these bins are highly predictive, thus contribute to the final classification results. The x-axis represents only the number of points, implying no ordering. The horizontal green line and red line are the mean values computed from green circles and red crosses, respectively. The bottom plot illustrates the generalization error from dividing brain voxels into 50 bins till 150 bins after collecting top informative features from the individual bins.
Table 3.
Workflow of proposed GMM+MS clustering method on 3D PET images.
Fig 3.
Number of features selected versus the classification accuracy using pSMC normalization.
Results are obtained by using the ADNI dataset on MCI against AD. Accuracy of BIC tend to drop after selecting 150 features, and AIC tends to drop after 500.
Fig 4.
Relation between the number of clusters and the number of bins on ADNI and TUM datasets.
Experiments are conducted on the normal control PET scans.
Table 4.
Result summary of three different methods on ADNI dataset.
Table 5.
Result summary of three different methods on TUM dataset.
Fig 5.
ROC curve of compared method on ADNI dataset.
To plot the curve, we collected the predicted probabilities for all the test sets in 10 times 10-fold cross-validation, along with their true class labels. For the plotting of ROC curves, we refer to page 173 of the book by Witten et al. [42].
Fig 6.
ROC curve of compared method on TUM dataset.
Table 6.
Top-10 informative regions (voxels) of MCI against AD using ADNI dataset using BIC.
Table 7.
Top-10 informative regions (voxels) of MCI against AD using TUM dataset using BIC.
Fig 7.
The informative regions (voxels) of MCI against AD using ADNI dataset of the 45th layer using BIC.
(a): coronal view (b): sagittal view (c): transaxial view. P: primary sensorimotor cortex normalization; G: grand mean normalization. x, y and z are the width (91), depth (109) and height (91) respectively. The red points represent the informative voxels, whereas other colors are only used to depict the brain structure.
Fig 8.
The informative regions (voxels) of MCI against AD using TUM dataset of the 45th layer using BIC.
(a): coronal view (b): sagittal view (c): transaxial view. P: primary sensorimotor cortex normalization; G: grand mean normalization. x, y and z are the width (91), depth (109) and height (91) respectively. The red points represent the informative voxels, whereas other colors are only used to depict the brain structure.