Abstract—Alzheimer’s disease (AD) is the most common form of dementia that causes progressive impairment of memory and cognitive functions of patients. However, whether imaging features can be utilised as biomarkers for this disease has not been explored. To address this, we encoded subcortical regions of the brain using 45 radiomic features to identify features specific for AD patients. We comprehensively evaluated the proposed approach using the OASIS dataset, assessing significance via the Wilcoxon test and Random Forest (RF) classifier models to identify the subcortical regions best able to identify AD patients. Our results show that features (i.e., correlation and volume) derived from several subcortical regions (i.e., cerebral, thalamus, caudate Putamen, Pallidum, hippocampus, amygdala, and stem-and-cerebrospinal-fluid) are able to identify AD from healthy control (HC) subjects with the hippocampus and amygdala reaching p < 0.01 following Holm-Bonferroni correction. Consistent with this, hippocampus (AUC = 81.19-84.09%) and amygdala (AUC = 79.70-80.27%) regions showed a higher AUC value compared to other subcortical regions. Combining radiomic features derived from all subcortical regions produced an AUC value of 91.54% for classifying AD from HC subjects. RF analysis revealed that from the 45 radiomic features, correlation and volume are the most important features for the classifier model. These results demonstrate that radiomic features extracted from hippocampus and amygdala regions are relevant biomarkers for AD patients and that correlation and volume features are the most important features to build this model.
Keywords—Alzheimer; radiomics; classification.
I. Introduction
Alzheimer disease (AD) is a progressive neurodegenerative disorder and the most common cause of dementia in the elderly [1]. Its symptoms consist of memory loss, confusion in time and place, lack of communication with others and failure to recognize family members [2]. As Alzheimer progresses, the brain cells die and connections among cell are lost which produces cognitive symptoms. This anatomo-pathology can be visualized non-invasively by MRI imaging, and described by radiomic analysis. For example, abnormal texture features are observed in the cerebellum white matter and hippocampus of patients with autistic spectrum disorder [3], [4].
Non-invasive quantitative imaging features could be a very promising technique to characterize the subcortical regions of AD patients. A variety of studies have been proposed to characterize and predict AD [5], with the most important biomarkers detected as brain volume and thickness. Villain et al., showed that the AD associated hippocampus is a third smaller than healthy patients [6]. A deep study of regional anatomical differences found in AD patients revealed that the mesial temporal region is the most effective region in the brain to identify AD (with mild cognitive impairment (MCI) ) [7].
In the last decade, many radiomic approaches based on texture features extraction has been widely considered for medical image analysis [8]–[12]. Radiomics in this context is the extraction of features from medical images (e.g., MRI) that are correlated to patient/disease characteristics. A novel study showed that constructing individual hierarchical networks of six 3D texture features from brain images indicates differences between tissues in AD, MCI, and HC [13]. Another study considered the combination of volume, cortical thickness, hippocampal shape and texture for differences diagnosis of HC, MCI, and AD; it was shown that the hippocampal texture was the most important feature in the algorithm followed by hippocampal volume, ventricular volume, and parietal lobe thickness [14]. However, limited studies considered the radiomics analysis to compare subcortical features for classifying between AD and HC subjects. In this context, the classification between AD and HC lead to better understanding of the neuroanatomical substrate of this disease.
In this study, we proposed to extract radiomic features for classifying the AD from HC subject using pre-labeled subcortical brain regions. Then using radiomic analysis, identify structure/texture regions which can best identify AD, thus providing an additional means of diagnosing AD and HC groups from structural MRI. This paper is structured as follows. Section 2 describes the techniques used to compute the radiomic features for classifying AD from HC subjects. Section 3 provides experimental setup and results. Sections 4 discusses our finding. Finally, Section 5 concludes with a summary of our work’s main contributions and results.
II. Materials and methods
Figure 1 shows this study’s workflow. Subcortical brain regions were pre-labeled automatically using FreeSurfer tool. Radiomic features were first computed for each of the brain subregions. A Wilcoxon rank sum test compared features between AD and HC subjects. We then considered all these features to train a RF model to discriminate between AD and HC subjects. This model was also used to identify the most discerning radiomic features.
Fig. 1. The proposed pipeline process of radiomic analysis for Alzheimer patients. 1) T1 weighted MR images acquired. 2) Automatic labeling of brain regions by FreeSurfer tool where each color represents subcortical region label. 3) Features are extracted from within the labeled subcortical regions, quantifying regions intensity, shape and texture. 4) Statistical analyses based on Wilcoxon test and RF models to identify relevant features of each brain region for classifying AD from HC subject.
This study included T1-weighted MRI data for 135 HC subjects with Clinical Dementia Rating (CDR) of 0, and 100 AD patients including 70 (CDR=0.5), 28 (CDR=1) and 2 (CDR=2) subjects were obtained from the publicly existing OASIS database (http://www.oasis-brains.org). Six preprocessing steps using the Freesurfer tool [15] were done upon T1 weighted MR images as follows: 1) small-motion correction by averaging the available volumes of subjects, 2) intensity normalization, 3) affine registration of volumes to the MNI305 atlas, 4) skull-stripping, 5) non-linear registration and further normalization using the Gaussian Classifier Atlas (GCA), and 6) brain parcellation and subcortical region labelling using the GCA. Gray-scale images were then normalized and acquired with a resolution of 1 mm3, for a total size of 256×256×256 voxels.
We sought identify the association between subcortical regions and variable structure of AD using radiomic features derived from brain MRI data. Using the pipeline Freesurfer tool [15], MRI volumes are first registered to an atlas and labelled into 25 key sub-cortical regions (i.e., left and right of the following subcortical brain regions: [Cerebral-White-Matter, Cerebellum-White Matter, Thalamus-Proper, Caudate, Putamen, Pallidum, Hippocampus, Amygdala, Accumbens-area, VentralDC and Vessel] and the Optic-Chiasm, Brain-Stem and cerebrospinal fluid (CSF) ). We then used all the extracted features (i.e., 45 radiomic features) as input for the RF model for classifying the AD from HC subjects based on each of subcortical regions.
A. Proposed radiomic features
We extracted over 1014 radiomic image features that could characterize the heterogeneity and shape of subcortical regions and simply computed. Specifically, we extracted a subset of 988 texture derived from 3D GLCM (19 functions × 13 angles × 4 offsets/displacement), 4 shapes, 6 intensities, 5 texture NGTDM and 11 texture GLSZM features. Computed radiomic features are described as below:
We summarized 988 GLCM texture features by averaging them in 19 features (i.e., angular second moment, contrast, correlation, sum of squares variance, homogeneity, sum average, sum variance, sum entropy, entropy, difference variance, difference entropy, information correlation 1, information correlation 2, autocorrelation, dissimilarity, cluster shape, cluster prominence, maximum probability and inverse difference [16]) that derived from computed GLCMs of image intensities uniformly resampled to 32 grey-levels. We then computed the five features based on NGTDM (i.e., coarseness, contrast, busyness, complexity and texture strength [17]), 11 features based on GLSZM ( i.e., small zone size emphasis, large zone size emphasis, low gray-level zone emphasis, high gray-level zone emphasis, small zone / low gray emphasis, small zone / high gray emphasis, large zone / low gray emphasis, large zone / high gray emphasis, gray-level non-uniformity, zone size non-uniformity and zone size percentage [18]). We summarized the texture feature in a total of 35 texture features that measure the heterogeneity of subcortical regions.
As we mentioned previously that we computed four geometrical features known as shape features (i.e., porosity, fraction dimension, surface-area and volume) that encode the morphological characteristics of subcortical regions, such as porosity [19], volume and surface area, that describe the geometrical changes over time in AD subjects [20]. Additionally, we summarized the intensity features by the average (i.e., mean), variance, skewness, kurtosis, energy and entropy that quantify the voxel intensity level distribution for brain subcortical regions. For example, high contrast subcortical regions have a high variance that represents the heterogeneity using the first order of statistical model. The total of 45 imaging features was considered in uni-multi variate analysis to identify the most significant subcortical regions between AD versus HC, and determine the most important features of this task.
B. Statistical analysis
We applied Wilcoxon test to identify subcortical brain regions where the texture features are statistically significant for the differences between AD and HC subjects. Note that the radiomic features are continuous variables. Since we compared 25 subcortical brain regions using 45 radiomic features that represent multiple comparisons of test (i.e., 25 sub-cortical regions × 45 texture features = 1125 tests), we corrected the p values obtained from significance test according to the Holm-Bonferroni method [21] in which we considered p < 0.01 as a significant sub-cortical regions to differentiate between AD and HC.
For multivariate analysis, the area under the receiver-operating characteristic curve (AUC-ROC) is considered to evaluate the performance of RF classifier model to classify between AD and HC subjects. Note that the AUC measures the entire two-dimensional area underneath the entire ROC curve. We chose to use RF (i.e., with 1000 trees) as is one of the most effective and general-purpose classification algorithms, running efficiently on large databases with thousands of input variable/features [22]. However, other classifiers could be used in this context. We then performed the 5-fold cross-validation strategy to obtained unbiased estimates of performance, where training images (i.e. features) are divided into 5 equal sized subsets and, in each fold, one subset is put aside for testing and the remaining 4 subsets are used to train the RF classifier. Finally, AUC value is computed as the average AUC obtained across the 5 folds.
To show the importance of features, we measured in each RF model and feature, the increase in prediction error resulting from the permutation of feature values across out-of-bag observations. The importance values computed for every RF tree and averaged over the entire ensemble. These values were then normalized by dividing them by the ensemble’s standard deviation. Finally, the importance of each of the features was obtained by averaging these normalized values across all 5 folds. Positive importance values indicate that feature is predictive, whereas negative importance values identify non-predictive features.
III. Results
Radiomic features differences related to AD:We first investigated radiomic differences between AD and HC patients. Figure 2 shows the heatmap of p values (in -log10 space) obtained from the Wilcoxon test. All the subcortical regions and radiomic features with corrected p values < 0.01 are marked with a black-green circle. We found that several radiomic features of the following subcortical brain regions are statistically different (corrected p < 0.01) between the AD from HC. These sub-cortical brain regions are the cerebral-white-matter, thalamus, caudate, putamen, pallidum, hippocampus, amygdala, accumbens, ventral DC and CSF. We observed a pronounced bilateral symmetry for the cerebral-white-matter, thalamus, caudate, pallidum, hippocampus, amygdala, and accumbens regions, as well as a notable asymmetry for putamen and ventral DC regions. Moreover, we found that the highest significance was achieved using GLCM-correlation, surface-area and volume features of hippocampal regions, following by amygdala regions. This shows the usefulness of several types of features as shape and texture features derived from subcortical regions of the brain to compare AD versus HC subjects.
Subcortical classification using the radiomic features:Having identified the radiomic features that can distinguish patients with AD from HC subjects, we then assessed each of the subcortical brain regions using multivariate analysis based on the RF model. Using all our radiomic features (n=45), the resultant AUC-ROC range values varied between 56.63-84.09%, dependent on subcortical brain regions (see Fig. 3). We observed that the right and left hippocampus regions achieved the highest AUC value of 81.19-84.09%, followed by the amygdala region (79.70-80.27%). We noticed that the subcortical brain regions (i.e., hippocampus and amygdala) that had the highest AUC values were identified previously to have the highest significance rate in previous analysis by Wilcoxon test. Additionally, combined features derived subcortical brain regions showed the best AUC value of 91.54% (see Fig 3).
Identification of the importance radiomic features: Figure 4 shows the heatmap of the importance value of features in each of subcortical brain regions for discriminating between AD and HC subjects. We found that the following features (i.e., Histogram (variance and entropy), GLCM (correlation, sum-variance and information correlation), NGTDM (coarseness and texture strength), GLSZM (large zone/ high gray emphasis and gray-level non-uniformity) and Shape (surface area and volume) have the greatest importance for identification of AD with the “correlation” and “volume” features in hippocampus and amygdala regions the most dominant. This result is consistent with our previous finding which identified these regions using Wilcoxon test and RF classifier model. In addition, we found that both texture (i.e. correlation) and shape (i.e., volume) features have the potential to discriminate AD from HC subjects.
Fig. 2.Heatmap of the p values (-log10 space) obtained from Wilcoxon test for radiomic features differences between AD (n=100) and HC (n=135) subjects. Black-green circles indicate subcortical brain regions showing significant radiomic feature with corrected p < 0.01 following Holm-Bonferroni.
Fig. 3. (Top) Bar graph of the area under the ROC curves for classifying AD vs HC subjects using the radiomic features for each of the subcortical brain regions. (Bottom) ROC curve of the combined features derived from subcortical brain regions for classifying AD from HC subjects.
Fig. 4. Heatmap of the importance value of features for classifying the AD from HC subjects based on each of the subcortical brain regions. Reported values correspond to the mean increase in prediction error obtained by permuting the values of individual features across out-of-bag observations [23].
IV. Discussion
Most of the current models for classifying AD from HC subjects are based on clinical (i.e., CDR and Mini-Mental State Examination score) or imaging volume features [24]. However, the radiomic features, which include histogram, texture, and shape features extracted from MR scans, provide a non-invasive means to predict AD in patients [25].
Our findings are consistent with previous studies, which found various texture and shape features to be strong biomarkers of AD [13], [25], [26]. This study provides a comprehensive set of radiomic analysis of subcortical brain regions and how they relate to AD diagnoses. More specifically, our results show the usefulness of each of radiomic features across different subcortical brain regions (see Fig. 2). In comparison to previous studies, our analysis investigated the link between subcortical brain regions and the radiomic features which were the most discriminative between AD and HC subjects (see Fig. 3). This approach is motivated by the hypothesis that radiomic features can encode the heterogeneity and capture the changes in subcortical brain regions within AD patients. Our experiments have demonstrated the usefulness of radiomic features extracted from subcortical brain regions to classify AD from HC subjects.
V. Conclusions
In this paper, we proposed to investigate radiomic features derived from different subcortical brain regions to identify AD, using a RF classifier model to identify the most important features. Our result suggests that the hippocampus and amygdala are the regions of the brain most different between AD and HC subjects and that “correlation” and “volume” are the most important features for diagnosing AD. Furthermore, our findings indicated that radiomic features including the histogram, texture and shape features can effectively predict AD patients.
References
[1] J. R. Petrella, R. E. Coleman, and P. M. Doraiswamy, “Neuroimaging and early diagnosis of Alzheimer disease: a look to the future,” Radiology, vol. 226, no. 2, pp. 315–336, Feb. 2003.
[2] “2014 Alzheimer’s disease facts and figures,” Alzheimers Dement., vol. 10, no. 2, pp. e47–e92, Mar. 2014.
[3] A. Chaddad, C. Desrosiers, and M. Toews, “Multi-scale radiomic analysis of sub-cortical regions in MRI related to autism, gender and age,” Sci. Rep., vol. 7, p. 45639, Mar. 2017.
[4] A. Chaddad, C. Desrosiers, L. Hassan, and C. Tanougast, “Hippocampus and amygdala radiomic biomarkers for the study of autism spectrum disorder,” BMC Neurosci., vol. 18, no. 1, p. 52, 2017.
[5] M. W. Weiner et al., “The Alzheimer’s Disease Neuroimaging Initiative: a review of papers published since its inception,” Alzheimers Dement. J. Alzheimers Assoc., vol. 8, no. 1 Suppl, pp. S1-68, Feb. 2012.
[6] N. Villain et al., “Relationships between hippocampal atrophy, white matter disruption, and gray matter hypometabolism in Alzheimer’s disease,” J. Neurosci. Off. J. Soc. Neurosci., vol. 28, no. 24, pp. 6174–6181, Jun. 2008.
[7] A. Qiu, C. Fennema-Notestine, A. M. Dale, M. I. Miller, and Alzheimer’s Disease Neuroimaging Initiative, “Regional shape abnormalities in mild cognitive impairment and Alzheimer’s disease,” NeuroImage, vol. 45, no. 3, pp. 656–661, Apr. 2009.
[8] P. Lambin et al., “Radiomics: extracting more information from medical images using advanced feature analysis,” Eur. J. Cancer Oxf. Engl. 1990, vol. 48, no. 4, pp. 441–446, Mar. 2012.
[9] A. Chaddad, M. J. Kucharczyk, and T. Niazi, “Multimodal Radiomic Features for the Predicting Gleason Score of Prostate Cancer,” Cancers, vol. 10, no. 8, Jul. 2018.
[10] A. Chaddad, P. Daniel, C. Desrosiers, M. Toews, and B. Abdulkarim, “Novel Radiomic Features based on Joint Intensity Matrices for Predicting Glioblastoma Patient Survival Time,” IEEE J. Biomed. Health Inform., pp. 1–1, 2018.
[11] A. Chaddad, S. Sabri, T. Niazi, and B. Abdulkarim, “Prediction of survival with multi-scale radiomic analysis in glioblastoma patients,” Med. Biol. Eng. Comput., pp. 1–14, Jun. 2018.
[12] A. Chaddad, P. Daniel, and T. Niazi, “Radiomics Evaluation of Histological Heterogeneity Using Multiscale Textures Derived From 3D Wavelet Transformation of Multispectral Images,” Front. Oncol., vol. 8, p. 96, 2018.
[13] J. Liu, J. Wang, B. Hu, F.-X. Wu, and Y. Pan, “Alzheimer’s Disease Classification Based on Individual Hierarchical Networks Constructed with 3D Texture Features,” IEEE Trans. NanoBioscience, 2017.
[14] L. Sørensen et al., “Differential diagnosis of mild cognitive impairment and Alzheimer’s disease using structural MRI cortical thickness, hippocampal shape, hippocampal texture, and volumetry,” NeuroImage Clin., vol. 13, pp. 470–482, 2017.
[15] B. Fischl, “FreeSurfer,” NeuroImage, vol. 62, no. 2, pp. 774–781, Aug. 2012.
[16] R. M. Haralick, “Statistical and structural approaches to texture,” Proc. IEEE, vol. 67, no. 5, pp. 786–804, May 1979.
[17] M. Amadasun and R. King, “Textural features corresponding to textural properties,” IEEE Trans. Syst. Man Cybern., vol. 19, no. 5, pp. 1264–1274, Sep. 1989.
[18] G. Thibault et al., “Texture indexes and gray level size zone matrix application to cell nuclei classification,” 2009.
[19] Z. Lewandowski and H. Beyenal, Fundamentals of Biofilm Research. CRC Press, 2007.
[20] N. Schuff et al., “MRI of hippocampal volume loss in early Alzheimer’s disease in relation to ApoE genotype and biomarkers,” Brain, vol. 132, no. 4, pp. 1067–1077, Apr. 2009.
[21] S. Holm, “A simple sequentially rejective multiple test procedure,” Scand. J. Stat., pp. 65–70, 1979.
[22] L. Breiman, “Random Forests,” Mach. Learn., vol. 45, no. 1, pp. 5–32, Oct. 2001.
[23] K. J. Archer and R. V. Kimes, “Empirical characterization of random forest variable importance measures,” Comput. Stat. Data Anal., vol. 52, no. 4, pp. 2249–2260, Jan. 2008.
[24] M.-J. Chiu et al., “Plasma tau as a window to the brain—negative associations with brain volume and memory function in mild cognitive impairment and early alzheimer’s disease,” Hum. Brain Mapp., vol. 35, no. 7, pp. 3132–3142, 2014.
[25] A. Chaddad, C. Desrosiers, and M. Toews, “Local discriminative characterization of MRI for Alzheimer’s disease,” in 13th IEEE International Symposium on Biomedical Imaging, ISBI 2016, Prague, Czech Republic, April 13-16, 2016, 2016, pp. 1–5.
[26] C. V. Dolph, M. Alam, Z. Shboul, M. D. Samad, and K. M. Iftekharuddin, “Deep learning of texture and structural features for multiclass Alzheimer’s disease classification,” in 2017 International Joint Conference on Neural Networks (IJCNN), 2017, pp. 2259–2266.
