Automated differential diagnosis of dementia syndromes using FDG PET and machine learning

Abstract

Background: Metabolic brain imaging with 2-[¹⁸F]fluoro-2-deoxy-D-glucose positron emission tomography (FDG PET) is a supportive diagnostic and differential diagnostic tool for neurodegenerative dementias. In the clinic, scans are usually visually interpreted. However, computer-aided approaches can improve diagnostic accuracy. We aimed to build two machine learning classifiers, based on two sets of FDG PET-derived features, for differential diagnosis of common dementia syndromes.

Methods: We analyzed FDG PET scans from three dementia cohorts [63 dementia due to Alzheimer's disease (AD), 79 dementia with Lewy bodies (DLB) and 23 frontotemporal dementia (FTD)], and 41 normal controls (NCs). Patients' clinical diagnosis at follow-up (25 ± 20 months after scanning) or cerebrospinal fluid biomarkers for Alzheimer's disease was considered a gold standard. FDG PET scans were first visually evaluated. Scans were pre-processed, and two sets of features extracted: (1) the expressions of previously identified metabolic brain patterns, and (2) the mean uptake value in 95 regions of interest (ROIs). Two multi-class support vector machine (SVM) classifiers were tested and their diagnostic performance assessed and compared to visual reading. Class-specific regional feature importance was assessed with Shapley Additive Explanations.

Results: Pattern- and ROI-based classifier achieved higher overall accuracy than expert readers (78% and 80% respectively, vs. 71%). Both SVM classifiers performed similarly to one another and to expert readers in AD (F1 = 0.74, 0.78, and 0.78) and DLB (F1 = 0.81, 0.81, and 0.78). SVM classifiers outperformed expert readers in FTD (F1 = 0.87, 0.83, and 0.63), but not in NC (F1 = 0.71, 0.75, and 0.92). Visualization of the SVM model showed bilateral temporal cortices and cerebellum to be the most important features for AD; occipital cortices, hippocampi and parahippocampi, amygdala, and middle temporal lobes for DLB; bilateral frontal cortices, middle and anterior cingulum for FTD; and bilateral angular gyri, pons, and vermis for NC.

Conclusion: Multi-class SVM classifiers based on the expression of characteristic metabolic brain patterns or ROI glucose uptake, performed better than experts in the differential diagnosis of common dementias using FDG PET scans. Experts performed better in the recognition of normal scans and a combined approach may yield optimal results in the clinical setting.

Keywords: FDG PET; dementia; differential diagnosis; machine learning; visual reading.

Figure 1

Flowchart of the general workflow. Top: 2-[¹⁸F]fluoro-2-deoxy-D-glucose positron emission tomography (FDG PET) scans from three dementia cohorts and healthy participants underwent pre-processing, and two set of features [pattern expression scores and standard uptake value ratio (SUVR) in 95 regions of interest (ROIs)] were extracted. Two multi-class support vector machine (SVM) models were built to classify scans either as dementia due to Alzheimer’s disease (AD), dementia with Lewy bodies (DLB), frontotemporal dementia (FTD) or normal. Most common label after 500 iterations was compared to gold standard, and performance metrics calculated. Bottom: FDG PET scans were visually evaluated by two expert readers. Scans could be read either as AD, DLB, FTD, normal or other (inconclusive or other neurodegenerative pattern). Labels were compared to gold standard and performance metrics were calculated. Clinical diagnosis at follow-up or cerebrospinal fluid (CSF) biomarker for Alzheimer’s disease was considered a gold standard. ADRP, Alzheimer’s disease-related pattern; DLBRP, dementia with Lewy bodies-related pattern; FTDRP, frontotemporal dementia-related pattern; DMN, default mode network pattern.

Figure 2

Confusion matrix of the classification results using visual reading. Expert readers correctly diagnosed 161 out of 206 cases and achieved 78% overall accuracy. Scans were read (Predicted Class) either as dementia due to Alzheimer’s disease (AD), dementia with Lewy bodies (DLB), frontotemporal dementia (FTD), normal (NC) or other (inconclusive (inc.) or other (oth.) neurodegenerative pattern). The label was compared to gold standard (True Class).

Figure 3

The classification results using a pattern-based support vector machine classifier. A classifier based on pattern expression values and multi-class support vector machine (SVM) correctly diagnosed 99 out of 127 cases and achieved 78% overall accuracy. (A) Confusion matrix of labels predicted by the SVM (Predicted Class) compared to gold standard (True Class). (B) One vs. all receiver operating characteristic (ROC) curves for the four possible labels. AD, dementia due to Alzheimer’s disease; DLB, dementia with Lewy bodies; FTD, frontotemporal dementia; NC, normal control; AUC, area under the curve; TPR, true positive rate; FPR, false positive rate.

Figure 4

The classification results using a region of interest-based support vector machine classifier. A classifier based on 95 regions of interest (ROIs) and multi-class support vector machine (SVM) correctly diagnosed 177 out of 206 cases and achieved 86% overall accuracy. (A) Confusion matrix of labels predicted by the SVM (Predicted Class) compared to gold standard (True Class). (B) One vs. all receiver operating characteristic (ROC) curves for the four possible labels. AD, dementia due to Alzheimer’s disease; DLB, dementia with Lewy bodies; FTD, frontotemporal dementia; NC, normal control; ROI, region of interest; AUC, area under the curve; TPR, true positive rate; FPR, false positive rate.

Figure 5

Flowchart depicting three different approaches to the explanation of the model. Initially, we used the entire dataset and assessed feature importance using neighborhood component analysis (NCA). The weights were used to rank the features from most to least important. NCA was then performed separately on just the training set for each iteration, and the retained features were plotted on a frequency histogram. To explain the support vector machine (SVM) model we employed Shapley Additive Explanations (SHAP). ROI, region of interest; SUVR, standard uptake value ratio; AAL, automated anatomic labeling atlas.

Figure 6

Feature selection frequency histogram of 40 most important features. Feature selection was performed on training set in each iteration using neighborhood component analysis (NCA), identifying the best non-zero lambda value corresponding to the minimum average loss. AAL, automated anatomic labeling; L, left; R, right.

Figure 7

The effect of the number of regions of interest included on the F1 scores. F1 scores reached their maximal values after including the 40 most important regions of interest (ROIs). AD, dementia due to Alzheimer’s disease; DLB, dementia with Lewy bodies; FTD, frontotemporal dementia; NC, normal controls.

Figure 8

Absolute average Shapley values for four groups plotted on an AAL template. The most important features for dementia due to Alzheimer’s disease (AD) classification included bilateral temporal cortices, cerebellum, bilateral lingual, and calcarine sulci, for dementia with Lewy bodies (DLB) occipital cortices, hippocampi and parahippocampi, amygdala, and middle temporal lobes, for frontotemporal dementia (FTD) bilateral frontal cortical regions, middle cingulum and anterior cingulum, and for normal controls (NC) bilateral angular gyri, pons, and vermis. Regions (features) are color-coded blue to red from least to most important.