Deformation Based Classification of Alzheimer’s Disease
Thomas Bonde Larsen1, Akshay Pai1, and Sune Darkner1

1Computer Science, University of Copenhagen, Copenhagen, Denmark

Synopsis

Effective and accurate diagnosis of Alzheimer’s disease (AD) purely based on structural magnetic resonance imaging (MRI) is a very pertinent clinical problem. We present a simple but highly accurate registration-based method to discriminate between the three classes of healthy controls (HC), mild cognitively impaired (MCI) and AD. The method uses the norm of the tangent space of the deformation in a K-nearest neighbor KNN classifier. The result show that for 60 subjects, 20 in each class using n-fold cross-validation an overall accuracy of 81.6% with 75% for HC, 85% MCI and 85% for AD.

Motivation

Effective and accurate diagnosis of Alzheimer’s disease (AD) purely based on structural magnetic resonance imaging (MRI) is a very pertinent clinical problem. Several attempts such as the recent CADDementia challenge have been made using primarily intensity-based features such as texture or other morphological features such as shape and volume. We present a simple but highly accurate registration-based method to discriminate between the three classes of healthy controls (HC), mild cognitively impaired (MCI) and AD. The method uses the norm of the tangent space of the deformation in a K-nearest neighbor KNN classifier. The result show that for 60 subjects, 20 in each class using n-fold cross-validation an overall accuracy of 81.6% with 75% for HC, 85% MCI and 85% for AD.

Material and Methods

A set of 60 1.5T MRI subjects from the ADNI database baseline, recently recommended by ADNI, was used. 20 random subjects from each group of Healthy Controls (HC), Mild Cognitive Impaired (MCI) and Alzheimer’s disease (AD) were selected. FreeSurfer's N3 corrections was used to bias correct the images and the brains where extracted using AFNI s 3D skullstrip. We used a symmetric affine registration, followed by a non-rigid registration using the symmetric version of the Discrete Diffeomorphic Deformations (D3) [1] registration framework a variation of the Stationary Velocity Field and Normalized Mutual Information using Locally Orderless Registration [2]. We computed the 2-norm of each of the velocity fields as the distance between two images. A vector containing the distances to all other images in the dataset was used to characterize each image in the set. We then use a weighted K-nearest neighbor and the Euclidian distance between the characterization vectors as a distance.

Experiments and Results

All 60 images where registered to all other images in the dataset. This process resulted in 1770 non-rigid pairwise registrations. As AD is characterized by a general atrophy to the brain we used 20x20x20 mm knot spacing in D3 with uniformly distributed evaluation points with a distance of 2x2x2mm. From the deformation fields each characterization vector, (the set distance to all images in the training set) was computed. We conducted 4 classification experiments HC vs. MCI vs. AD, HC vs. MCI, HC vs. AD and MCI vs. AD based on the characterization vectors. The accuracy of the classification was tested using n-fold cross validation (leave one out) and the k number of neighbors in the KNN-classifier was set to k=10 (half the number of sample in each class).

The results of the pairwise classification can be found in Figure 1. Over all the classification accuracy ranges from 80-90%. For the 3 class problem the overall classification accuracy is 81.6%. A detailing of the results is found in Figure 2.

Discussion

The results are very promising and indicate that the distance defined by image registration, in this case the norm of the SVF, to the rest of the population is a remarkable descriptor for the classification of AD. The results achieved are roughly 30% better what have previously been achieved. The results also indicate that it may be beneficial and provide more balanced results if all 3 classes are considered simultaneously. If this methodology generalizes to the entire ADNI database and others is impossible to say, but the results are on their own remarkable. Furthermore, the 3 class problem indicate that a ordering in some non-linear subspace may exist.

Conclusion

We have shown that using D3 registration tool at a semi coarse scale combined with a KNN-classifier we can discriminate between HC, MCI and AD. The average results of both the 2-class has an accuracy of 80-90% and that 3 class problem has a classification accuracy of 81.6% with 75% for HC, 85% MCI and 85% for AD. These results vindicate the utility of image registration in not only longitudinal analysis but also cross-sectional analysis of brain MRI.

Acknowledgements

No acknowledgement found.

References

1 S. Darkner, A Pai, M.G Liptrot and J Sporring "D3: Discrete Diffeomorphic Deformations for Image Registration", (major revision) Neuro Image

2 S. Darkner and J Sporring "Locally Orderless Registration" IEEE Transactions on Pattern Analysis and Machine Intelligence 06/2013; 35(6):1437-1450.

Figures

The 2-class classification results. Blue is the true class label and green is the outcome of the classifier. As the figure shows, the classification between AD and HC is comparable to state of the where as the MCI vs AD and HC vs MCI produces remarkably good results.

This figure show the results of the 3-class classification task. Blue is the true class label and green is the outcome of the classifier. The results are well balanced between the classes. The overall classification accuracy is 81.6%. Notice how AD is only misclassified as MCI and HC as MCI.



Proc. Intl. Soc. Mag. Reson. Med. 24 (2016)
4047