Since the world population is aging rapidly, the prevalence of dementia is also rising rapidly thus causing a great impact on individuals, families and societies. Accurate classification and level measurement of dementia are very importance in the disease management. Numerous studies show that 18F-FDG-brain scan can differentiate various types of dementia. However, correct and accurate interpretation of nuclear images requires physicians who are well experienced. Therefore, it is worthwhile to build an automatic diagnostic system for it. In this paper, we present a novel method by using an artificial neural network (ANN) to analyze CortexID of brain PET-CT scan with clinical and laboratory data for dementia classification. Moreover, the ANN was trained to indicate the clinical severity of the disease as reflected by MMSE score. All ANNs were trained and tested again with an experienced physician’s seventy diagnosis and the results were very promising. The dementia classifier achieved 96% accuracy and the mapper network could correctly predict the MMSE score with 0.782 regression value.
The Hong Kong population is aging rapidly such that the elderly population (those aged 65 years or older) is estimated to increase from approximately 13% in 2009 to 28% by 2039, there will be 2.5 million elder people in Hong Kong [
Dementia is a group of diseases that primarily affect older people; but it is not a normal part of aging. The World Health Organization’s International Classification of Disease (1992-1994) defines dementia as:
“Dementia is a syndrome due to disease of the brain, usually of a chronic or progressive nature, in which there is disturbance of multiple higher cortical functions, including memory, thinking, orientation, comprehension, calculation, learning capacity, language and judgment. The impairments of cognitive function are commonly accompanied, and occasionally preceded by deterioration in emotional control, social behavior, or motivation”.
There are numerous different forms of dementia and each has its own causes. Dementia can also be a symptom that develops in the later stages of other illnesses. The common types and causes of dementia include:
・ Alzheimer’s disease (AD) [
・ Vascular dementia (VaD) [
・ Mixed dementia (MD) [
・ Lewy bodies dementia (LBD) [
・ Frontotemporal Lobar Degeneration (FTLD) [
・ Mild Cognitive Impairment (MCI) [9 , 10] is considered as a prodromal stage of dementia, with noticeable cognitive changes, but no significant disability to interfere with daily life.
There is no single diagnostic test for all dementias and a comprehensive assessment (such as clinical history taking, cognitive and physical examination, medication review, brain scanning, etc.) is required in order to exclude treatable causes of cognitive impairment.
Common assessments include:-
・ Mini-mental state examination (MMSE), using in initial checkup, to analyze a subject’s cognitive status. It uses a set of questions to provide measures of orientation, registration, short-term memory, motor and language functioning.
・ Routine heametology and biochemistry tests to detect untreated syphilis, abnormality of thyroid function and decline of B12.
・ MRI/CT brain scan to show the level of any atrophy and any vascular disease or infarcts.
However, absolute confirmation of the diagnosis can only be made by a biopsy and an autopsy of brain tissues.
Numerous studies have found that AD and other dementias could alter brain metabolisim significantly. AD patients usually present the posterior cingulate and parietotemporal cortices hypometabolisim and spread into the frontal lobes in advanced stage. In contrast, FTLD patients show manifestly hypometabolism in the frontal and anterior temporal cortices, while DLB patients present hypometabolism in the posterior brain comprising primarily the parietoocipital regions. Theoretically, 18F-FDG PET scan can help in early diagnosis of AD and other dementias by highlighting these decreased FDG uptake cortex regions before MRI or CT scans can detect any structural brain damage.
To quantify the cortex metabolic activity, the reconstructed FDG PET images of the patient are processed with CortexID, developed by Dr. Satoshi Minoshima [12 , 13]. The software provides automated brain assessment through quantification of the comparison of local peak activity values, at standardized anatomical locations, with the corresponding reference normal peak activity in age-stratified control subjects. 3D stereotactic surface projections of the brain are generated and 31 regional Z-scores are calculated. A positive Z-score is given in a hypometabolic CortexID area.
By using these 31 regional Z-scores with MRI cerebrovascular condition, two ANNs were designed for dementias differentiation and MMSE score mapping. The details will be presented in section 4.
ANN is a mathematical model to simulate the structure and functionalities of biological neural networks. The basic building block, artificial neuron, consists of multiplying every input value with individual weight, summing all weighted values and activating through the activation function (
y ( k ) = g ( ∑ i = 1 m w i ( k ) ⋅ x i ( k ) + b ) (1)
where x i ( k ) , w i ( k ) and y ( k ) are input, weight and output values in discrete time k respectively,
b is bias,
g is a activation function.
A multilayer perceptron (
Therefore MLP is widely and successfully used for pattern classification.
MLP utilizes the backpropagation rule for training. The rule repetitively calculates an error function for each input and backpropagates the error from one layer to the previous one. The weights are adjusted in direct proportional to the error in the nodes to which it is connected. The back-propagation algorithm is described below:
Step:
1) Initialize the weights randomly.
2) Arbitrarily select an input X.
3) Feed the signal through the network.
4) Compute δ i L in the output layer
δ i L = g ′ ( h i L ) ( d i − y i L ) (2)
where h i L is the net input to the ith neuron in the Lth layer,
g ′ is the derivative of the activation function,
d i and y i L are the desired and feed forward output respectively.
5) Back-propagate the errors and compute the preceding layers deltas
δ i l = g ′ ( h i l ) ∑ j w i j l + 1 δ j l + 1 , l = ( L − 1 ) , ⋯ , 1 (3)
6) Update weights
Δ w i j l = η δ i l y i l where η ∈ ( 0 , 1 ) (4)
7) Repeat step 2 until the error in the output layer is below a prespecified values.
MLPs are universal function approximators so they can be used to create mathematical models for function mapping as well as classifier. Therefore, MLPs were chosen in our application. Since too much hidden layers results in memorizing and not learning, one hidden layer is sufficient for the tasks. Except linear output neurons were used in MMSE score mapping, all neurons employ sigmoid activition function and trained with back-propagation algorithm.
108 PET-CT examination radiological reports from University of Hong Kong at Queen Mary Hospital were screened. 87 reports are dementia cases and 17 of them were excluded because of incomplete CortexID data or unclassified or only having a single incidence. Moreover, 55 reports have MMSE scores and the frequency distribution of dementia is tabulated in
It is observed that the distribution is severe skewed which will provide poor accuracies in most ANNs. To cope with the problems, class distributions need to be normalized by creating synthetic samples with Z-scores perturbed randomly [
| Type of Dementia | Number of cases |
|---|---|
| MCI | 13 |
| AD | 12 |
| VaD | 12 |
| MD | 27 |
| LBD | 6 |
We used Matlab 2010 neural toolbox to build the ANNs. The data was divided randomly into the training set, the validation set and the test set. The training set is used for computing the gradient and updating the network weights. The validation set is used to improve generalization and to stop training when the network starts to overfit the data. The test set error is for different models comparison. For the dementia classifier training, the data is partitioned with a 60% training set, 20% validation set and 20% test set. In MMSE score mapping, the division portions were adjusted to 70%, 10% and 20% respectively.
In order to determine the optimum number of hidden neurons, the amount of neurons in the hidden layer is progressive increasing until the errors of the training, validation and test data sets are below the prespecified threshold.
The obtained NN models for dementia classifier (
The performance of the classifier can be visualized in the confusion matrices (
are very close to the upper left corner, this implies the overall accuracy of the classifier is very high.
The correlations between the outputs of mapping function network and desired MMSE target values were plotted in
To prevent biased, untrained data sets are used to calculate the sensitivities and specificities of the dementia classifier. Other related performance indices are also computed and tabulated in
Based on the analysis table, it is observed that the dementia classifier was learnt to categorise different types of dementia very accurately as a radiologist by using 31 regional CortexID Z-scores with cerebrovascular condition.
In MMSE score mapping study, the sample sizes of validation and testing data sets are 6 and 11; their regression values are 0.983 and 0.646. Therefore, the combined regression value is
| Dementia | True Positive | False Positive | True Negative | False Negative | Sensitivity | Specificity | PPV | NPV | Accuracy |
|---|---|---|---|---|---|---|---|---|---|
| MCI | 10 | 2 | 42 | 0 | 100% | 95.5% | 83.3% | 100% | 96.3% |
| AD | 9 | 0 | 43 | 2 | 81.8% | 100% | 100% | 95.6% | 96.3% |
| VaD | 12 | 0 | 40 | 2 | 85.7% | 100% | 100% | 95.2% | 96.3% |
| MD | 8 | 2 | 44 | 0 | 100% | 95.7% | 80% | 100% | 96.3% |
| DLB | 11 | 0 | 43 | 0 | 100% | 100% | 100% | 100% | 100% |
| Dementia | Type I Error | Type II Error | Area under ROC curve (AUC) | ||||||
| MCI | 4.5% | 0% | 0.985 | ||||||
| AD | 0% | 18.2% | 0.979 | ||||||
| VaD | 0% | 14.3% | 0.981 | ||||||
| MD | 4.3% | 0% | 0.985 | ||||||
| DLB | 0% | 0% | 1 | ||||||
R validation + testing = 6 × 0.983 2 + 11 × 0.646 2 6 + 11 = 0.782
Since the R validation + testing is very close to 1, the mapping network correlates highly with MMSE score.
Two multilayer perceptrons, based on 31 regional CortexID Z-scores, were designed and trained to classify different types of dementia and map with MMSE score. Comparing with physicians diagnosis, the dementia classifier achieved 96% accuracy. The mapper network can correctly predict the MMSE score with 0.782 regression value.
However, due to the limit number of records, we do not know whether the network can be trained to differentiate other types of dementias, such as frontotemporal dementia, Parkinsonian syndromes, etc. Further study and verification are required.
This is a retrospective chart review study and is supported by the Department of Diagnostic Radiology, University of Hong Kong and the Hospital Authority, Hong Kong. We also wish to acknowledge the clinical data support of Prof. Leung-Wing Chu of the Department of Medicine of the University of Hong Kong.
The authors declare no conflicts of interest regarding the publication of this paper.