Filters and Segmentation
As mentioned in the materials and methods section Jerman and Frangi filters were compared to apply the best enhancement filter to the dataset. For each filtering method different combination of filtering values were also tested to acquire the most efficient parameter values. Based on the quantitative and qualitative results the best filter was Frangi with the following parameter values:
𝛾 = 0.07, S_min=1.5
The dice score acquired with these parameters was the highest value (dice = 0.749).
The following tables show the final dice score values for both Jerman and Frangi filters which were acquired by averaging over all the eight datasets. For each filtering method the highest dice score values are bolded.
Table.1: Averaged dice score values using Frangi filter with different filtering parameter values
Table.2: Averaged dice score values using Jerman filter with different filtering parameter values
In addition to the quantitative comparison, the qualitative method also showed better performance of Frangi compared to Jerman filter. In figure 1 you can see the behaviour of both Frangi and Jerman filters on the same dataset. The performance of filters is compared with each other with respect to the ground truth which was obtained semi-manually using ilastik.
(a) Ground truth
(b) Frangi enhanced (c) Jerman enhanced
Fig.1: Comparison between Frangi and Jerman filters with respect to the ground truth
Statistical Analysis
As mentioned, 12 brain regions are compared between healthy and AD subjects. One of the main results obtained after calculating the metrics in each of these brain regions was the low vascular depiction. This low vessel depiction is due to the MRI resolution (3T) and the filtering method. As seen in the previous section, the highest dice value was 0.749 which shows that the filtering method is not perfect.
Hippocampus was the only region in which all the subjects (24 healthy and 24 AD) had detectable vascular information. In thalamus for around half of the subjects, vessels could be detected. In the remaining ROI either few or no subjects had detectable vascular information. It should be noted that even though the vessels could be detected in the hippocampus of all subjects the vascular depiction was still below 1%. Based on these results one could conclude that due to the low vessel depiction, the vascular metrics are not generally very sensitive in capturing any probable vascular alterations associated with AD.
Based on Mann-Whittney U-test, no significant difference was found in the vessel volume and vessel density metrics between healthy and AD subjects. Figure 2 shows the boxplots of the vessel volume values in both healthy and AD subjects.
Fig.2: Comparison between the vessel volume of healthy and AD subjects in six brain regions based on box plots
Even though U-test did not show any significant difference in the vessel volume between healthy and AD subjects, the ESM values (ESM = 0.28) revealed a moderate vessel volume loss in the left hippocampus of people suffering from AD.
In contrary to vessel volume and vessel density, the vascular alterations could be captured by VDM. Based on the U-test and ESM, a significant difference was found in the left hippocampus between healthy and AD category (P = 0.04, ESM = -0.445). Figure 3 shows the VDM boxplots.
Fig.3: Comparison between the VDM of healthy and AD groups in six brain regions based on the boxplots
It can be easily seen that more similarity exists between the left and right hippocampus of the healthy subjects compared to the AD group. The values of VDM in the left and right hippocampus of AD patients tend to be more asymmetrical and have higher standard deviations compared to the healthy subjects (higher interquartile range). The significant difference in the VDM values between healthy and AD categories can be easily seen in the box plots.
Discussion
We saw that despite low vessel depiction VDM was able to capture AD-related vascular alterations in the left hippocampus which could not be detected by other vascular metrics.
The results that we obtained supported the hypothesis of the presence of reverse correlation between vessel density/volume and VDM (moderate loss of vessel volume in the left hippocampus of AD patients resulted in an increase in the VDM values).
It can be concluded that VDM seems to be a promising approach for detecting vascular alterations. However, due to some limitations including MRI resolution and filtering approach further investigation is required to approve the reliability of this approach.
Comments