Abstract
Objective
Recent research indicates that using multi-modal data, such as PET/CT, MRI/EEG and DTI/fMRI, could achieve better performance than using single-modal data. To quantitatively analyze the synergy between multi-modal biomarkers, we proposed a novel cross-view pattern analysis method in this study.
Methods
We investigated two types of neuroimaging biomarkers derived from the T1-weighted MRI and the FDG-PET data of 331 ADNI subjects. We first identified the disease-relevant patterns of each biomarker by applying ANOVA on the regions of interest across the whole brain, and then quantified the synergy between these two biomarkers based on the relative entropy of their patterns. To validate the effectiveness of proposed method, we designed a set of experiments to distinguish patients with Alzheimer’s Disease (AD) or Mild Cognitive Impairment (MCI) from the cognitively normal controls (NC). We further tested the null hypothesis of no correlation against the alternative that there was a nonzero correlation between the true classification performance and the predicted synergy.
Results
The patterns of these two biomarkers varied dramatically. The PET biomarker showed a more spreading pattern than the MRI biomarker and achieved higher precision for NC (53.3%) and MCI subjects (64.0%), whereas MRI biomarker had better performance on AD patient classification (67.6%). Combining the multi-modal biomarkers could achieve better results (63.9% for NC, 64.5% for MCI and 80.6% for AD) as compared to either PET or MRI biomarker. Cross-validation showed that the performance improvements were closely correlated to the predicted synergy.
Conclusion
We proposed a cross-view neuroimaging pattern analysis method to quantify the synergy in multi-modal data. The pilot experiment results showed that our method could reliably predict the synergy between different biomarkers.
Significance
Multi-modal data are playing a more important role in medical image analysis, but it is challenging to take full advantage of them. Our cross-view pattern analysis method showed a great potential to evaluate the synergy between different modalities and predict their performance.
Recent research indicates that using multi-modal data, such as PET/CT, MRI/EEG and DTI/fMRI, could achieve better performance than using single-modal data. To quantitatively analyze the synergy between multi-modal biomarkers, we proposed a novel cross-view pattern analysis method in this study.
Methods
We investigated two types of neuroimaging biomarkers derived from the T1-weighted MRI and the FDG-PET data of 331 ADNI subjects. We first identified the disease-relevant patterns of each biomarker by applying ANOVA on the regions of interest across the whole brain, and then quantified the synergy between these two biomarkers based on the relative entropy of their patterns. To validate the effectiveness of proposed method, we designed a set of experiments to distinguish patients with Alzheimer’s Disease (AD) or Mild Cognitive Impairment (MCI) from the cognitively normal controls (NC). We further tested the null hypothesis of no correlation against the alternative that there was a nonzero correlation between the true classification performance and the predicted synergy.
Results
The patterns of these two biomarkers varied dramatically. The PET biomarker showed a more spreading pattern than the MRI biomarker and achieved higher precision for NC (53.3%) and MCI subjects (64.0%), whereas MRI biomarker had better performance on AD patient classification (67.6%). Combining the multi-modal biomarkers could achieve better results (63.9% for NC, 64.5% for MCI and 80.6% for AD) as compared to either PET or MRI biomarker. Cross-validation showed that the performance improvements were closely correlated to the predicted synergy.
Conclusion
We proposed a cross-view neuroimaging pattern analysis method to quantify the synergy in multi-modal data. The pilot experiment results showed that our method could reliably predict the synergy between different biomarkers.
Significance
Multi-modal data are playing a more important role in medical image analysis, but it is challenging to take full advantage of them. Our cross-view pattern analysis method showed a great potential to evaluate the synergy between different modalities and predict their performance.
| Original language | English |
|---|---|
| Number of pages | 1 |
| Publication status | Published - 13 Nov 2014 |
| Externally published | Yes |
| Event | IEEE EMBS BRAIN Grand Challenges Conference - Washington DC, United States Duration: 13 Nov 2014 → 14 Nov 2014 |
Conference
| Conference | IEEE EMBS BRAIN Grand Challenges Conference |
|---|---|
| Country/Territory | United States |
| City | Washington DC |
| Period | 13/11/14 → 14/11/14 |