A differential diagnosis system for classifying demented patients based on their MRI scans is presented in this paper. The main findings of the paper are: 1) Patterns identified in the Differential STAND maps in each neurodegenerative dementia (AD, FTLD-TDP and LBD) are unique and mirror anatomic patterns of pathological neurodegeneration established in autopsy studies. 2) Differential classification based on structural MRI scan provides reasonable differential diagnostic accuracy of dementia in subjects antemortem with accuracy of differentially diagnosing AD 87%, LBD 95% and FTLD 90%. Importantly we describe a proof-of-concept differential diagnostic system based on structural MRI that considers the three common neurodegenerative dementias simultaneously as opposed to the more typical approach of systems that distinguish a single dementing disorder from cognitively normal elderly subjects.
An important feature of the proposed system is the use of “pure” autopsy confirmed cases for building the differential diagnosis system. This addresses two potentially significant confounding issues; situations where clinical presentation does not predict the correct underlying pathology, and patients with mixed pathologies which would obfuscate the desired objective of mapping MRI patterns onto specific neurodegenerative pathologies. Other studies have established regions of atrophy based on MRI for differential diagnosis between AD and FTD (Davatzikos et al., 2008
; Foster et al., 2007
; Horn et al., 2009
; Kloppel et al., 2008a
; Rabinovici et al., 2007
). Likewise, (IMP) SPECT (Ishii et al., 2009
) and (11C-DTBZ and 18F-FDG) PET studies (Koeppe et al., 2005
) have addressed differential diagnosis of FTD, AD and DLB in clinical subjects. All of these studies have focused on pair wise differential diagnosis (i.e. AD vs. FTD, AD vs. DLB etc) but none have addressed the simultaneous differential diagnosis of all three common neurodegenerative causes of dementia using patterns of neurodegeneration from antemortem scans of pathology confirmed dementia cases.
In AD, the topographic distribution of neurofibrillary tangles (NFT) follows a fairly stereotypical pattern of progression (Braak and Braak, 1991
) and it has been shown that atrophy seen on MRI correlates well with pathological Braak NFT stages (Gosche et al., 2002
; Jack et al., 2002
; Vemuri et al., 2008b
; Whitwell et al., 2008
) and with NFT density (Csernansky et al., 2004
; Silbert et al., 2003
). The neurodegenerative patterns we found in pathologically confirmed AD, as illustrated in , are concordant with the Braak NFT pattern.
Pathologically, FTLD is very heterogeneous (Forman et al., 2006
; Hodges et al., 2004
; Josephs et al., 2006a
) and can be characterized by the deposition of both TDP-43 and tau. In this paper we considered only subjects with FTLD-TDP, the most common pathology underlying the frontotemporal dementias, because we wanted to develop a differential diagnosis system for diagnosis of pathology rather than diagnosis of clinical dementia type. Our results concur with previous pathological and imaging studies that have demonstrated progressive degeneration of the frontal and temporal lobes, with relative sparing of the parietal and occipital lobes, in frontotemporal dementia (Broe et al., 2003
; Kril et al., 2005
; Seeley et al., 2008
; Whitwell et al., 2009b
). Patterns of atrophy typically vary according to the specific clinical syndrome (Rosen et al., 2002
), although we have allowed for this variability by applying a cluster-based approach to classification.
In LBD, the major pathological findings are Lewy bodies and Lewy neurites with degeneration of several neurotransmitter systems, most notably dopaminergic and cholinergic. The concentration of LB pathology is high in the amygdalae and various brain stem locations with low densities in the neocortical regions (Dickson, 2002
; Klucken et al., 2003
). A recent pathology study demonstrated that LBD is associated with depletion of cholinergic neurons in the pedunculopontine tegmental and laterodorsal tegmental nuclei present in the dorsal midbrain (Schmeichel et al., 2008
) which agrees with the findings of a recent voxel based morphometry (VBM) study of DLB patients (Whitwell et al., 2007
). The most severe supratentorial LB pathology is observed in the amygdale, often early in the disease (Marui et al., 2002
). Also dysfunction of the visuo-amygdaloid pathway has been implicated in visual misidentification and visual hallucinations (Yamamoto et al., 2006
). LBs and degeneration of amygdala might be one of the underlying causes of visual misidentification in LBD patients (Harding et al., 2002
; Iseki et al., 2001
; Yamamoto et al., 2006
). The inferior temporal lobe is a unimodal visual association cortex related to the visual recognition network (Benarroch, 2006
) and the structural abnormality of inferior longitudinal fasciculus has also been observed in diffusion tensor studies in DLB (Kantarci et al., 2010
; Ota et al., 2009
)., Thus, the regions of GM loss we found in LBD patients () are consistent with the imaging and pathology literature in LBD.
One of the key problems in the field of dementia diagnosis is the antemortem separation of LBD subjects from AD. LBD subjects are often misclassified as AD and the sensitivity of DLB diagnosis in different studies varies between 0–100 percent (McKeith et al., 2004
). AD and LBD pathology both increase in prevalence dramatically with age. The pathological hallmarks of AD (NFTs and neuritic plaques) and LBD (LBs) often coexist and there is evidence that the same neuronal circuits are disturbed within the hippocampal formation in both diseases (Klucken et al., 2003
). For this reason there is a high probability of including mixed AD and LBD cases in MRI studies that are based on clinical diagnosis, which often leads to the conclusion that the patterns of atrophy overlap in AD and LBD. A unique feature of this study was that we included only pathologically confirmed LBD cases which helped us identify the following distinctive features that differentiate LBD from AD: 1) significant gray matter loss is confined to the amygdalae and middle temporal lobe in LBD whereas in AD it is throughout the entire medial temporal lobe (Gomez-Isla et al., 1999
; Hashimoto et al., 1998
; Lippa et al., 1998
) and 2) decreased GM is present in the dorsal ponto-mesencepalic junction area in LBD (Schmeichel et al., 2008
). Even though the sensitivity of LBD diagnosis was only 79%, the high specificity of 99 % indicates that it can be used to detect “pure” LBD cases with a relatively small number of false positives. In the specific example shown in , MRI features were able to classify all ten patients in cluster #4 correctly as LBD where as three out of the ten LBD subjects were misclassified as AD clinically. This cluster consists of mild LBD cases which might be difficult to separate clinically from AD early in the disease process in the absence of all the LBD symptoms. Another interesting observation was cluster #3 where mild AD and LBD patients were clustered together due to similarity of features.
An MRI based automated differential diagnosis technique such as the one proposed here could be a very useful adjunct tool to clinical evaluation because it requires minimal human intervention, adjusts for aging related changes and may be able to extract subtle changes in brain structures that are difficult to assess by visually examining MRI scans. The proposed classification approach based on clustering has several advantages: Diagnosis is only made on stable clusters which may either be “pure” clusters (only a single dementia type) or “mixed” clusters (clusters that are a mixture of two or more dementia types) as shown in . The arbitrary clusters are excluded from the decision making process due to the lack of a clear cluster structure based on the features from MRI. In the cases of “pure” clusters, Differential-STAND based pathology diagnosis has the highest chance of being correct while diagnosis on clinical information could possibly be misleading in a few cases. In the “mixed” clusters, the confidence of the MRI diagnosis is lower than that of the clinical diagnosis due to the considerable overlap in the MRI features. Theoretically the end user could be cautioned if the new incoming patient falls in a mixed cluster in order to facilitate an informed decision.
Additionally, Differential-STAND Maps provide a visual assessment of the ROIs that appear to be abnormal when compared to a normal database and when evaluated by the differential diagnosis system, provide information about the likely type of dementia pathology in demented patients. The construction of the algorithm using multi-ROI level GM volume rather than voxel-level GM volume addresses two issues: 1) smaller number of features containing GM sampled from across the entire brain rather than a very high dimensional feature set of GM voxels which will not provide good differential diagnosis due to the limited number of training samples and 2) reduces the number of false positives of voxel level differences that arise due to registration and segmentation errors.
Even though there are several two-class classification frameworks that are typically applied in the neuroimaging literature, the approach presented in this paper provides a conceptually elegant solution for the multi-class problem. The main reasons being 1) moving forward, the clustering of subjects into dementia sub-types might provide useful clinical or pathological subtype information regarding the disease; 2) the proposed approach is extremely useful in determining the pathology underlying mixed dementias which constitute of more than 40 % of all dementia cases. If we wanted to use a two-class SVM based framework like we used in our previous papers, then we would have to make a decision of applying either a pair-wise SVM or one-against-all SVM which is not conducive while attempting to visualize the distance of a specific patient from the different dementia sub-types. In contrast, the proposed approach naturally lends itself to simultaneously determining where the subject falls with respect to various dementia sub-types.
Value of Differential STAND approach to clinical diagnosis
There can be considerable mismatch between the antemortem clinical diagnosis and the gold standard postmortem pathology, specifically in FTLD-TDP and LBD. Autopsy confirmed studies have found that most FTD patients also fulfill the diagnostic criteria of AD (Varma et al., 1999
) and LBD subjects are often misclassified as AD (McKeith et al., 2004
). There can be low inter-rater reliability for differential diagnosis of neurodegenerative dementias, with the lowest generalized kappa of 0.37 for DLB diagnosis in (Lopez et al., 1999
). Additionally, there can be considerable syndromic heterogeneity, e.g. some pathology confirmed FTLD-TDP cases have prominent anterograde amnesia while some pathology confirmed AD cases have mainly aphasia (Alladi et al., 2007
; Josephs et al., 2008
). There were some pathological confirmed cases which highlight this syndromic heterogeneity where both clinical as well as differential STAND classification were incorrect and both methods made similar mistakes possibly due to atypical patterns of atrophy presenting atypical clinical symptoms. Given that proteinopathy does not map exactly onto the clinical expression of the disease, we believe that complementary measures, such as the output of this system, can add great value to clinical diagnosis in conjunction with established clinical evaluation methods.
We acknowledge that the diagnostic accuracy of the Differential-STAND system is very similar to the accuracy based on clinical classification. However, we envision the added value of Differential-STAND system to clinical diagnosis will occur in subjects who have dementia that is difficult to classify (which the Differential-STAND approach was able to make correct pathological classification in all the cases), mixed dementia pathologies, mildly symptomatic subjects early in the disease process who have not yet presented symptoms clinically (e.g. mild LBDs in cluster #9), and clinically atypical cases. It has in fact been previously shown at the group-level that patterns of atrophy on MRI can help identify the presence of AD pathology in subjects with atypical clinical syndromes (Josephs et al., 2010
; Whitwell et al., 2009a
), suggesting that an individual-level differential-STAND approach could be very useful. Finally we note that while the diagnostic performance of Differential-STAND was similar to that of clinicians, the clinicians who proved the diagnoses in our study were highly skilled behavioral Neurology specialists practicing in a large tertiary referral medical center (Mayo Clinic). We expect that diagnostic performance of clinicians who do not see large numbers of neurodegenerative dementia patients would be worse. Diagnostic performance data support this assumption, for example the sensitivity of DLB diagnosis varies between 0–100 percent (McKeith et al., 2004
) depending on the expertise of the clinician. In contrast, Differential-STAND is not based on clinical expertise but simply on a subject’s MRI scan which is independent of site to site variation in clinical expertise. We therefore expect that a major clinical utility of the Differential-STAND system would be in situations where highly skilled clinical experts are not available. Similar findings have been reported in the FDG PET literature indicating that the use of imaging (PET scans) can add valuable information to clinical evaluation particularly in situations where highly skilled specialists are not available (Silverman et al., 2002
Limitations of this study
- A scheme for pathological staging of disease severity that maps to cortical atrophy exists for AD (Braak staging) but analogous pathological schemes for staging severity and atrophy do not exist for FTLD-TDP or LBD. If accepted pathological severity staging schemes existed for FTLD-TDP and LBD, it would allow the development and validation of a system that could assign the relative contribution of each pathology subtype. However at the present time, the proposed system would deal with the problem of mixed pathologies by assigning the subjects with mixed pathologies to the atrophy pattern matching the Differential-STAND Maps of the dominant neurodegenerative pathology (underlying the dementia).
- Strength of this study is the use of autopsy diagnoses as the gold standard. This is at the same time a weakness, because to our knowledge, no center world wide including our own is in possession of a large enough sample of subjects with antemortem protocol MRIs and autopsy diagnoses in all three diseases to permit splitting the sample into independent training and test samples. Due to the small number of autopsy confirmed pure dementia cases, we do not have an independent test dataset where new subjects would be input into the Differential-STAND Map based system that had not been used to construct the classifier..
- It is inevitable that the identified pathologically confirmed CN subjects are older than the dementia cases specifically in the FTLD-TDP group. However the application of the age-adjustment of GM based on a uniform distribution of a large cohort of CN subjects ensures that this age effect is minimized.
- Vascular pathology has not been considered in this work at the present time and will be part of our future training databases.