Skip to main content
NCBI home page
NeuroImage : Clinical logoLink to NeuroImage : Clinical
. 2023 Apr 17;38:103405. doi: 10.1016/j.nicl.2023.103405

Explainable classification of Parkinson’s disease using deep learning trained on a large multi-center database of T1-weighted MRI datasets

Milton Camacho a,b,⁎, Matthias Wilms b,c,1, Pauline Mouches a,b,1, Hannes Almgren d,e,1, Raissa Souza a,b, Richard Camicioli f, Zahinoor Ismail d,e,g, Oury Monchi d,e,h,i, Nils D Forkert b,c,d,e,j
PMCID: PMC10148079  PMID: 37079936

Highlights

  • •

    3D CNN was trained to detect Parkinson's disease only using structural brain images.

  • •

    Jacobian maps help generalizability when training from a large multi-center dataset.

  • •

    Explainable artificial intelligence leads to understandable predictions.

Keywords: Magnetic Resonance Imaging, Deep learning, Parkinson’s disease, Explainable artificial intelligence

Abstract

Introduction

Parkinson’s disease (PD) is a severe neurodegenerative disease that affects millions of people. Early diagnosis is important to facilitate prompt interventions to slow down disease progression. However, accurate PD diagnosis can be challenging, especially in the early disease stages. The aim of this work was to develop and evaluate a robust explainable deep learning model for PD classification trained from one of the largest collections of T1-weighted magnetic resonance imaging datasets.

Materials and Methods

A total of 2,041 T1-weighted MRI datasets from 13 different studies were collected, including 1,024 datasets from PD patients and 1,017 datasets from age- and sex-matched healthy controls (HC). The datasets were skull stripped, resampled to isotropic resolution, bias field corrected, and non-linearly registered to the MNI PD25 atlas. The Jacobian maps derived from the deformation fields together with basic clinical parameters were used to train a state-of-the-art convolutional neural network (CNN) to classify PD and HC subjects. Saliency maps were generated to display the brain regions contributing the most to the classification task as a means of explainable artificial intelligence.

Results

The CNN model was trained using an 85%/5%/10% train/validation/test split stratified by diagnosis, sex, and study. The model achieved an accuracy of 79.3%, precision of 80.2%, specificity of 81.3%, sensitivity of 77.7%, and AUC-ROC of 0.87 on the test set while performing similarly on an independent test set. Saliency maps computed for the test set data highlighted frontotemporal regions, the orbital-frontal cortex, and multiple deep gray matter structures as most important.

Conclusion

The developed CNN model, trained on a large heterogenous database, was able to differentiate PD patients from HC subjects with high accuracy with clinically feasible classification explanations. Future research should aim to investigate the combination of multiple imaging modalities with deep learning and on validating these results in a prospective trial as a clinical decision support system.

1. Introduction

Parkinson’s disease (PD) is a severe and heterogeneous progressive disease that affects millions of people. It is recognized as the second-most common neurodegenerative disease affecting 2–3% of the population aged 65 years or older (Poewe et al., 2017). Pathophysiologically, PD is characterized by the loss of dopaminergic neurons in the substantia nigra and intracellular alphasynuclein accumulation that can spread in various regions of the cortex as the disease progresses (Braak et al., 2003, Fearnley and Lees, 1991). PD is diagnosed following clinical guidelines for assessing the presence of bradykinesia and at least one cardinal motor symptom, which includes tremor, rigidity, and postural instability (Balestrino and Schapira, 2020). Other motor and non-motor symptoms complement the clinical presentation (Balestrino and Schapira, 2020). However, previous studies have found that diagnostic accuracy based on clinical guidelines can vary from 73.8% to 83.9% depending on neurologist experience (Rizzo et al., 2016). Therefore, ongoing research aims to identify highly accurate and reproducible biomarkers for PD diagnosis. In particular, a considerable body of research has investigated biomarkers extracted from neuroimaging such as magnetic resonance imaging (MRI) and positron emission tomography for this purpose due to their high spatial resolution and high sensitivity to display the structure and function of the brain. For example, as the brain ages, there is a natural degree of atrophy, which can be considered normal. Nevertheless, neurodegenerative diseases like PD can cause this degeneration to accelerate exponentially (Baecker et al., 2021). T1-weighted MRI allows detecting early morphological changes of the brain that are promising for the diagnosis of PD (Ibarretxe-Bilbao et al., 2012, Sarasso et al., 2021, Talai et al., 2021, Tessa et al., 2014, Tsiouris et al., 2022). However, it can be challenging to visually identify the subtle differences associated with PD in these high-resolution datasets.

Supervised machine learning methods aim to automatically identify relevant patterns in a high-dimensional feature space based on historic datasets and corresponding ground truth classification and use these uncovered patterns to classify previously unseen cases (lo Vercio et al., 2020). Based on this, machine learning models have been developed that aim to make use of such high-dimensional data for individual-level classification (Maceachern and Forkert, 2021). Within this context, a variety of machine learning models for PD classification have been developed based on structural neuroimaging with accuracy levels ranging from 71.5% to 100% (Adeli et al., 2017, Adeli et al., 2016, Amoroso et al., 2018, Chakraborty et al., 2020, Cigdem et al., 2018, Esmaeilzadeh et al., 2018, Solana-Lavalle and Rosas-Romero, 2021). Most of these works followed a classical machine learning setup, typically including image preprocessing, feature extraction, and training and testing of a conventional machine learning model (e.g., support vector machine) using a cross-validation scheme. One major limitation of these methods is that the features need to be explicitly defined beforehand, potentially missing important features. Convolutional neural networks (CNN) are a unique type of artificial neural networks that are specifically designed for the automatic analysis of image data and are capable to optimize the feature extraction from images by using convolutional filters as well as the subsequent classification. A few previous works have proposed CNN models for PD classification (Chakraborty et al., 2020, Esmaeilzadeh et al., 2018) reporting a much higher accuracy of 92.75% to 100% when compared to conventional machine learning models. However, these models typically need a large amount of data for the training process and often overfit the data, especially when trained on single-center datasets.

To date, previous works have only included a limited and often imbalanced number of PD patients and healthy participants ranging from 40 to 374 subjects, in many cases acquired within a single study or even just in a single center. In contrast, in our study we are including data from 13 different studies with over five times as many subjects. Moreover, most previously proposed machine learning models do not provide an explanation for the classification. Explainable classifications with high accuracy represent an important step towards increasing the trust of patients and caregivers in the computer-aided diagnosis system. This is clearly a lack in current literature that our study addresses. Thus, the aim of this work was to collect a large sample of structural T1-weighted MRI datasets from PD patients and matched healthy controls acquired in multiple studies to train a robust and accurate CNN model for classification that can explain its decision by highlighting the regions that were most influential for classification of an individual patient.

2. Materials and methods

2.1. Datasets and subjects

Data were obtained from 13 different studies comprising 1,051 PD patients and 1,026 control subjects (2,077 participants in total). The studies included were the Parkinson’s Progression Markers Initiative (PPMI)2, Canadian Consortium on Neurodegeneration in Aging (CCNA) COMPASS-ND (Duchesne et al., 2019), BioCog (Clinical, Magnetic Resonance, and Genetic Biomarkers of Cognitive Decline and Dementia in Parkinson’s Disease) (Acharya et al., 2007), PD-MCI Calgary (Lang et al., 2019), PD-MCI Montreal (Hanganu et al., 2014) (Longitudinal Study on Mild Cognitive Impairment in Parkinson’s Disease), C-BIG3 (Montreal Neurological Institute’s Open Science Clinical Biological Imaging and Genetic Repository), NEUROCON dataset4 (Badea et al., 2017), Tao Wu dataset4 (Badea et al., 2017), OpenNeuro Japan dataset5, Hamburg dataset (Boelmans et al., 2012), United Kingdom Biobank (UK Biobank)6, OASIS3 (Longitudinal Neuroimaging, Clinical, and Cognitive Dataset for Normal Aging and Alzheimer Disease) (LaMontagne et al., 2018), and Southwest University Adult Lifespan Dataset (SALD) (Wei et al., 2017). For this secondary study, we included all participants that met the following criteria: (1) Diagnosis of either PD or healthy control according to each individual study, (2) availability of T1-weighted MRI datasets (earliest available in case of longitudinal studies like PPMI and BioCog), and (3) preprocessed scans passing our quality control checks (Section 2.2). Our final subject count after preprocessing and quality control included 2,041 subjects (1,024 PD patients and 1,017 healthy controls [HC]). Table 1 contains demographic, clinical, and diagnosis criteria information of the participants used for the present work. Table S1 contains information regarding the inclusion and exclusion criteria of the different studies. Each study received ethics approval from their local ethics board and received written informed consent from all the participants in accordance with the declaration of Helsinki. In addition, the present study was approved by the Conjoint Health Research Ethics Board at the University of Calgary.

Table 1.

Demographic information.

Study Total Subjects
(female %)
 Age in Years
Mean (std)
 Duration of Disease in Years
Mean (std)		 H&Y
Mean (std)
 MoCA
Mean (std)
 Vendor
Name(N of scans acquired)		 Clinical Diagnostic Criteria
PD HC PD HC PD PD HC PD HC
PPMI 539 (36.36%) 166
(36.74%)		 62.35
(9.57)		 60.14
(11.65)		 1.44 (1.76) 1.63 (0.51) 0.006
(0.07)		 26.90 (2.51) 28.18 (1.29) S(405) GE(176) P
(124)		 MDS
COMPASS-ND 59
(33.33%)		 – 68.64
(7.50)		 – – 2.01
(0.63)		 – 24.67
(4.68)		 – S(55) GE(4) MDS
BioCog 45
(44.44%)		 49
(42.85%)		 70.31
(3.81)		 71.28
(4.96)		 4.00
(5.21)		 – – – – S(94) UKBB
PD-MCI Calgary 79
(32.91%)		 42
(52.38%)		 71.31
(6.41)		 69.80
(6.87)		 6.04
(4.55)		 – – 25.50 (3.37) 27.35
(1.94)		 GE(121) MDS
C-Big 66
(45.45%)		 10
(90%)		 65
(8.43)		 62.4
(11.88)		 – – – – – S(76) MDS
NEUROCON 26
(37.03%)		 16
(75%)		 68.76
(10.7)		 67.62
(11.88)		 – 1.92
(0.33)		 0.0
(0.0)		 – – S(42) MDS
Tao Wu 17
(52.9%)		 20
(40%)		 64.5
(4.19)		 64.75
(5.58)		 4.8
(3.46)		 1.91
(0.64)		 0.0
(0.0)		 – – S(37) MDS
ON Japan 30
(56.66%)		 15
(53.33%)		 67.56
(6.80)		 63.33
(5.24)		 – – – – – S(45) UKBB
Hamburg 74
(29.7%)		 – 63.62
(8.95)		 – 12.33
(6.17)		 2.52
(0.76)		 – – – S(74) UKBB
UK Biobank 48
(41.66%)		 197
(39.59%)		 70.02
(9.91)		 (70.02)
(7.76)		 7.48 (6.61) – – – – S(245) UKBB
OASIS3 – 403
(37.4%)		 – 66.51
(7.26)		 – – – – – S(403) –
SALD – 78
(0.0%)		 – 61.65
(8.03)		 – – – – – S(78) –
PD-MCI Montreal (Independent test set) 41
(36.58%)		 21
(52.38%)		 61.81
(5.75)		 62.85
(6.36)		 5.02
(3.82)		 – – 27.34
(1.99)		 27.90
(1.94)		 S(62) UKBB
Total 1,024
(37.69%)		 1,017
(37.56%)		 64.74
(9.18)		 65.23
(8.75)		 7.37
(6.09)		 2.41
(0.77)		 0.001
(0.04)		 26.58
(2.93)		 28.01
(1.51)		 S(1616) GE(301) P(124) MDS and UKBB
Open in a new tab

H&Y = Hoehn and Yahr rating scale; MoCA = Montreal Cognitive Assessment test; MDS = Movement Disorder Society clinical diagnosis criteria; UKBB = United Kingdom Parkinson’s Disease Society Brain Bank clinical diagnostic criteria; S = Siemens; GE = General Electric; P = Philips.

2.2. MRI acquisition

T1-weighted MRI images were acquired using a variety of scanners and sequence parameters across the 13 studies. Siemens (for 12 studies), GE (for 3 studies), and Phillips (for 1 study) were among the MRI scanner manufacturers with magnetic fields of either 3.0 T (11 studies with 1669 subjects) or 1.5 T (4 studies with 374 subjects). The slice and in-slice resolution ranged from 0.94 to 2.0 mm, and most of the images were acquired in the sagittal plane. The supplementary material (Table S2) provides more details regarding the imaging protocols.

2.3. Preprocessing

T1-weighted images were preprocessed executing the following steps. First, the original MRI scans underwent brain extraction using HD-BET (Isensee et al., 2019) to remove all non-brain tissue from the images. The resulting images were then resampled to an isotropic resolution of 1 mm with linear interpolation. After that, we performed bias field correction with the non-parametric non-uniform intensity normalization algorithm from the Advanced Normalization Tools (ANTs) toolkit version 2.3.1 (Tustison et al., 2010). Next, each T1-weighted MRI dataset was non-linearly registered to the Montreal Neurological Institute (MNI) PD25-T1-MPRAGE-1 mm brain atlas (fixed image) (Xiao et al., 2017). This was performed using ANTs by first aligning the datasets to the atlas using an affine transformation, which was then used for initialization of the non-linear registration performed in a second step (Avants et al., 2011). All registration results were visually checked to identify any misaligned datasets. The displacement fields resulting from the non-linear registration were used to compute the corresponding log-Jacobian maps using ANTs (Leow et al., 2007). Briefly explained the log-Jacobians reflect local differences in volume between the atlas and the individual datasets for each voxel and are symmetric around 0. Thus, values below zero represent a volume decrease, and values above zero represent a volume increase, which makes them better suited for training of neural networks than standard Jacobian maps. The Log-Jacobian maps were masked using the MNI PD25 brain mask to remove undesired volume changes outside the brain (background voxels) induced by ANTs’ regularizer. The larger the input images, the higher the model complexity is and more computational requirements are needed to optimize CNN models. Therefore, the images were cropped to match the defined dimensions for the inputs of the CNN models (the CNN models implemented were defined to accept 3D images of 160x192x160 voxels), to remove unnecessary background voxels, and to reduce computational burden during training (Fig. 1).

Fig. 1.

Fig. 1

Open in a new tab

Preprocessing pipeline. The diagram illustrates the preprocessing of the datasets resulting in affinely registered (bottom left), non-linearly registered (bottom center), and Jacobian maps (bottom right). QC = quality control.

2.4. Extraction of volumes of interest from raw images

For comparison purposes, the MNI ICBM 152 atlas was registered to the MNI PD25 atlas using the previously described non-linear registration step to obtain the Harvard-Oxford (HO) cortical and subcortical segmentations in the MNI PD25 atlas space. Afterwards, the inverse non-linear transformation for each patient was used to warp the HO cortical and sub-cortical atlas (Desikan et al., 2006, Frazier et al., 2005, Goldstein et al., 2007, Makris et al., 2006) brain regions to each individual T1-weighted MRI dataset employing a nearest-neighbor interpolation. The HO subcortical atlas includes 21 brain areas such as the putamen, caudate, hippocampus, and brainstem, whereas the HO cortical atlas includes 48 brain regions like the insular cortex, orbital cortex, and temporal pole (Fig. 2). The volumes of the included cortical/sub-cortical anatomical regions were computed and normalized using the intracranial volume (each structure volume/intracranial volume) calculated from every patient’s brain mask previously obtained with HD-BET. These computed volumes were later utilized to train a conventional machine learning model to use as a baseline comparison to the proposed deep learning architecture.

Fig. 2.

Fig. 2

Open in a new tab

Pipeline for extraction of cortical and subcortical volumes. On the left, the Harvard-Oxford atlases overlayed on top of the ICBM 152 atlas are shown while the transformed Harvard-Oxford segmentations overlayed on the PD25 atlas after registration are shown on the right. After warping the Harvard-Oxford atlases to the subject space, the volumes are computed including the calculation of the intracranial volume derived from the previous brain extraction.

2.5. Deep learning and machine learning models

The deep learning model used in this work was adapted from the simple fully convolutional neural (SFCN) network, which achieved state-of-the-art performance in adult brain age prediction and sex classification using T1-weighted MRI datasets (Peng et al., 2021, Stanley et al., 2022). Moreover, SFCN has successfully been used in combination with methods to compute saliency maps for diverse applications (Mouches et al., 2022, Stanley et al., 2022). Briefly described, the architecture used for this work consisted of a 3D input layer of dimension 160x192x160, five convolutional blocks, each containing a 3D convolutional layer with a 3 × 3 × 3 kernel, batch normalization, 2 × 2 × 2 max pooling (except for the fourth block), and ReLU activation. A sixth convolutional block consisted of a 3D convolutional layer with a 1 × 1 × 1 kernel, a batch normalization layer, a 3D average pooling layer, and ReLU activation. The convolutional filter sizes for each individual block were 32, 64, 128, 256, 256, and 64. A seventh block consisted of a dropout layer with a rate of 0.5, with the output flattened and passed to a single classification node with sigmoid activation. All convolutional layers included L1 regularization (Tibshirani, 1996) with a 0.0005 regularization factor.

A single study (PD-MCI Montreal) was held out as an independent test set to evaluate the best performing classification model. The rest of the data was split into 85% (n = 1,682) training, 5% (n = 98) validation, and 10% (n = 199) testing data with an equal percentage of participants from each group (PD vs. HC), sex, and data origin included in every split. Briefly explained, this was done to avoid introducing data biases during training and to help counter the data imbalances in the individual studies. The model was trained from scratch using the log-Jacobian maps (CNNJacobians) with randomly initialized weights for 180 epochs with an early stopping criterion determined by the model’s performance based on the validation data. During training, a binary cross entropy loss was optimized using the Adam optimizer (Kingma and Ba, 2017). The loss was weighted during model training such that equal importance is given to each of the studies with respect to data size differences and patient distribution. More precisely, these weights were defined as 1 minus the proportion of subjects from a given group (PD or HC) that each study contributed to the total number of datasets on that group (e.g., PPMI provided 539 PD subjects so that the weight was = 1 − (539/1,024)). Furthermore, the learning rate was empirically set to 0.003 with learning decay of 0.0007 and a batch size of 5.

Three other models were created for comparison purposes. Previous studies have primarily used preprocessed intensity images to train their models rather than log-Jacobian maps (Chakraborty et al., 2020, Esmaeilzadeh et al., 2018). Thus, two additional models were trained including intensity information. First, the affinely registered intensity images were used to train the CNN model described above (CNNIntensity) without regularization, which was found to produce better results for this model. After that, a second CNN model was trained combining both affinely registered intensity datasets and Jacobian maps (CNNCombined) as the input of the model (creating a two-channel input with the following dimensions 160x192x160x2). These two models were trained using the same data splits, architecture, and validation techniques used for CNNJacobians. Although the settings were nearly identical, the baseline models were optimized with different learning rates: 0.005 for CNNIntensity and 0.0001 for CNNCombined, which was found beneficial for the performance of the models.

The last classification model was developed using a Nu support vector machine (Nu-SVM) with a radial basis kernel function (RBF) (El-Manzalawy and Honavar, 2005) in addition to the CNNs for a comparison with previous literature employing SVMs (Adeli et al., 2017, Amoroso et al., 2018, Cigdem et al., 2018, Solana-Lavalle and Rosas-Romero, 2021). This model was trained and optimized using the previously extracted volumes from the Harvard-Oxford atlases (see section 2.3) with the same splits described before, a Nu parameter of 0.5, and gamma parameter of 0.03. To identify the most predictive brain volumes and limit the number of redundant and non-informative features, RELIEFF feature selection with 10 nearest neighbours (Kononenko et al., 1997) was utilized to rank the features based on their importance. An iterative process was performed removing the least important feature and re-running the model, creating a new model each time a feature was dropped until only two features remained. RELIEFF was selected for this purpose because it can detect non-linear relationships between features and the output class. The model within the Nu-SVM-RBF framework with the best performance was selected to serve as the additional baseline representing the combination of feature engineering with classical machine learning models.

Finally, the models (comparing all of them) that performed the best for the test split data were evaluated by producing predictions for a completely independent test set (PD-MCI Montreal).

2.6. Saliency map calculation

To identify what regions contributed most for the classification between HC and PD, saliency maps were generated for the best performing CNN model. These were produced using the SmoothGrad algorithm (Smilkov et al., 2017) for every successfully classified participant of the test set using CNNJacobian, which performed best overall for the test set and significantly better for the independent test set (Section 3). Briefly explained, this method adds random noise from a Gaussian distribution to each test dataset and feeds the noisy images into the trained model. The method then computes and backpropagates the corresponding partial derivative of the loss function with respect to the noisy input image. As a result, one value is assigned to each image voxel to reflect the voxel's importance to the model's output. This process is performed 25 times to smoothen the saliency maps with a noise standard deviation of 0.01. The saliency maps’ intensities are then min–max normalized to a range from 0 to 1, and the resulting maps are averaged. Because the Log-Jacobians images are already in the same atlas space, no additional post-processing was required to spatially normalize the saliency maps.

A single saliency map was created from the successfully classified test datasets by averaging the saliency maps across all test subjects to gain a better understanding of the model's general behaviour. The average map's lower voxel intensity values were removed by applying a threshold of 0.5 to focus on the most important regions being highlighted.

2.7. Identification of regions of importance from deep learning

The Harvard-Oxford cortical and subcortical atlas brain regions (see section 2.3) were used to quantitatively analyze the computed average saliency maps locally and to identify the most important regions for the classification healthy subjects and patients with Parkinson's disease. For this purpose, the percentage of each Harvard-Oxford atlas region covered by the thresholded average saliency map, as well as the mean saliency map value, were calculated and investigated.

2.8. Evaluation metrics

The metrics used to evaluate the performance of the models included accuracy, precision, specificity, sensitivity, and area under the receiver operating characteristic curve (ROC-AUC), which measures model performance while comprehensively accounting for the trade-off between true positive rate and false positive rate (lo Vercio et al., 2020). Furthermore, a paired t-test was used to determine statistical significance between the models’ predictions where a p-value < 0.05 was considered to indicate statistical significance.

3. Results

Table 2 illustrates the prediction metrics for the CNNJacobians, CNNIntensity, CNNCombined, and SVM-RBF Parkinson’s classification models in terms of the accuracy, precision, specificity, sensitivity, and ROC-AUC. The proposed CNN model using Log-Jacobians (CNNJacobians) achieved the best overall results for the split test set (79.3% accuracy, 80.2% precision, 81.3% specificity, 77.7% sensitivity, and 0.87 ROC-AUC) when compared to the other models. Based on the performance for the split test set, the CNNIntensity and the CNNCombined models did not perform significantly worse than the best performing model (Log-Jacobians) but SVM-RBF did. Therefore, the CNNJacobians, CNNIntensity, and CNNCombined models were evaluated on the independent test set to identify the best model to produce saliency maps. Here, the CNNJacobians model achieved significantly better performance (p < 0.05) than the other two models (intensity and combined). When the CNNJacobians model was applied to the independent test set, the accuracy, specificity, and ROC-AUC dropped, while the precision and sensitivity increased (77.4% accuracy, 81.3% precision, 61.9% specificity, 85.3% sensitivity, and 0.86 ROC-AUC) when compared to the results for the non-independent test set.

Table 2.

Performance of the trained models for classification of healthy subjects and patients with Parkinson’s disease.

Model Test set ACC PRC SPE SEN ROC-AUC
CNNJacobians Split 79.3% 80.2% 81.3% 77.7% 0.87
Independent 77.4% 81.3% 61.9% 85.3% 0.86



CNNIntensity Split 78.8% 78.7% 79% 78.7% 0.87
Independent + 61.2% 67.3% 23% 80% 0.74



CNNCombined Split 71.3% 73.8% 77% 65.6% 0.79
Independent + 62.9% 72.5% 47.6% 70.7% 0.65



Nu-SVM-RBF Split* 66.8% 66.3% 66% 67.7% 0.67
Open in a new tab

ACC = accuracy, PRC = precision, SEN = sensitivity, SPE = specificity, ROC-AUC = area under the receiver operating characteristic curve.

* = significant difference (p < 0.05) with the best model for the split test set data.

+ = significant difference (p < 0.05) with the best model for the independent test set data.

Fig. 3 shows the saliency map derived from CNNJacobians classifications for the correctly diagnosed PD patients in the split test set. The regions highlighted are the brain areas used by the model to make classifications whereas greater saliency map values indicated an increased importance for the classification task.

Fig. 3.

Fig. 3

Open in a new tab

Average saliency maps of the correctly classified PD patients overlayed on the PD25 atlas. Panel A shows three coronal slices from posterior to anterior, panel B shows three axial slices from inferior to superior, and panel C shows three sagittal slices from left to right. L = left; R = right; A = anterior; P = posterior.

Fig. 4 depicts the degree to which the saliency maps cover the Harvard-Oxford brain atlas sub-cortical and cortical regions, together with the average and standard deviation of the saliency values per region. The brain regions identified as most important by the CNNJacobians model included subcortical regions like the amygdala, putamen, pallidum, hippocampus, nucleus accumbens, and cortical regions including the orbitofrontal, insular cortex, parahippocampal gyrus, temporal fusiform cortex, planum polare, occipital fusiform gyrus, temporal pole, subcallosal cortex, and paracingulate gyrus. For the SVM machine learning model, the most important features selected by the RELIEFF feature selection method are shown in Table S3. Generally, the brain regions selected by the traditional machine learning method are similar to the results of the saliency maps computed from the CNNJacobians model and included the amygdala, putamen, hippocampus, pallidum, nucleus accumbens, parahippocampal gyrus, orbitofrontal and medial frontal cortex, temporal pole, temporal fusiform cortex, occipital fusiform gyrus, subcallosal cortex, planum polare, and paracingulate gyrus.

Fig. 4.

Fig. 4

Open in a new tab

Volume percentage of the Harvard-Oxford atlases regions covered by the thresholded average saliency map of correctly classified PD patients. Additionally, every region has an mean (standard deviation) intensity value for the corresponding volume percentage covered, also indicated by the color of the corresponding bar. Regions were ranked according to the mean intensity value of the saliency map.

4. Discussion

The major contribution of this work is the development of a robust and explainable deep learning model for the automatic differentiation of patients with PD from healthy subjects. This model was trained and evaluated on one of the largest databases containing T1-weighed MRI datasets from 13 different studies that used a wide range of MRI scanners and sequence parameters and was additionally tested on a completely independent database not used for training or testing.

4.1. Parkinson’s disease detection

In the present study, the deep learning models achieved considerably better results compared to the conventional machine learning approach (Nu-SVM-RBF). At the same time, the CNN using the Jacobian maps led to a considerably better accuracy when compared to CNNs using either the intensity information from the affinely registered T1-weighted MRI datasets alone or a combination of the affinely registered T1-weighted MRI datasets and Jacobian maps together. The finding that deep learning outperforms conventional machine learning tasks is expected and in line with the current literature showing improved performance of deep learning models compared to conventional machine learning models for many tasks (Adeli et al., 2017, Adeli et al., 2016, Aishwarya and Ravi Kumar, 2021, Amoroso et al., 2018, Chakraborty et al., 2020, Cigdem et al., 2018, Esmaeilzadeh et al., 2018, Sharma et al., 2021, Shinde and Shah, 2018, Solana-Lavalle and Rosas-Romero, 2021). This is partly explained by the end-to-end learning nature of the deep learning method implemented as opposed to feature engineering where important information might be missed in the process. Moreover, the potential reason for the improved performance of the CNNJacobians model compared to the CNNIntensity and CNNCombined models is likely related to the considerable variation of T1-weighted images used in this study. Essentially, the Jacobian maps provide a more homogeneous representation of the morphological information in the T1-weighted images acquired on many different scanners using different imaging parameters. However, more elaborate intensity normalization techniques may also help to boost the performance of the CNN models using the intensity information from the affinely registered T1-weighted datasets. To the best of our knowledge, this is the first study investigating the use of 3D Jacobian maps to train a CNN for PD diagnosis.

Another benefit of the proposed deep learning architecture is the rather simple CNN architecture used, which contains only 2,959,939 trainable parameters. While complex architectures can achieve better results, they are often more prone to overfitting issues, resulting in poor performance on unseen data. Compared to larger CNNs like the ones presented by Chakraborty et al. (2020) using 16.7 million trainable parameters and Esmaeilzadeh et al. (2018) using 69.7 million trainable parameters, the proposed model’s simplicity supports the model generalizability. Aside from the technical simplicity, the proposed model only requires a preprocessed T1-weighted MRI dataset of the patient in question (which are widely available in clinical settings) so that an implementation of this model within a computer-aided diagnosis system in the clinical setting is feasible. In this context, our model could act as a decision support system that clinical experts could use to obtain additional valuable information for the diagnosis, which could lead to improved PD diagnosis accuracy in clinical practice.

The prediction accuracy of the CNNJacobians model on unseen data is within the range of what has been previously described in literature based on T1-weighted datasets (71.5% to 100% accuracy) (Adeli et al., 2017, Adeli et al., 2016, Amoroso et al., 2018, Chakraborty et al., 2020, Cigdem et al., 2018, Esmaeilzadeh et al., 2018, Solana-Lavalle and Rosas-Romero, 2021). Nonetheless, the machine learning models used in this study were trained on a much larger and balanced database of PD patients and HC subjects showing large variations regarding the imaging data available. The results achieved on the test data (which is also composed of datasets collected from 12 different studies) and a completely independent study not included in the training and test sets show that the suggested model generalizes well to new unseen data but loses some accuracy when applied to the completely independent dataset, which is expected. The data used for this study generally resembles that of real clinical brain imaging data. Thus, the developed classification model represents a step towards the goal of creating a computer-aided PD diagnosis tool capable of identifying the subtle brain atrophy changes associated with the disease regardless of the origin or imaging system used for acquisition. In this context, the proposed model could also be further refined as more data becomes available. Nevertheless, the model developed in this work needs to be tested in a prospective trial as a diagnosis support tool for new patients with suspected PD at early stages of the disease where accurate diagnosis can be most challenging. The advantages of this approach are two-fold. First, an early prediction could facilitate a more accurate diagnosis and, second, a later diagnosis confirmation could be used to validate the models’ performance at the early phase.

4.2. Important brain regions

To the best of our knowledge, this is the first work to apply methods from the explainable artificial intelligence domain to unbox “black box” CNN models for PD classification from Jacobian maps, which is needed to enhance trust in this machine learning model in clinical practice, for patients and caregivers at the same time. Through the use of explainable artificial intelligence methods, the developed model could help clinical experts to evaluate the model predictions’ feasibility and utilize the produced saliency maps for biomarker discovery. Generally, the brain regions identified in this work are in line with our current understanding of this disease and the affected brain regions.

More precisely, six subcortical regions and eleven cortical regions were identified by the CNNJacobians model as being most important. The subcortical regions included the bilateral amygdala, bilateral pallidum, bilateral hippocampus, bilateral putamen, bilateral nucleus accumbens, and caudate (mostly left hemisphere). In line with these findings, a large previous multicenter study identified volumetric changes in the bilateral putamen, left amygdala, bilateral hippocampus, left globus pallidus, right nucleus accumbens, and left caudate in PD patients compared to healthy controls (Laansma et al., 2021). These brain regions also exhibited high mean saliency map values (importance) in the results of this work.

The cortical regions identified as being most important by the model included the frontal cortex (medial, subcallosal, orbital, and operculum cortex), insular cortex, temporal pole, parahippocampal gyrus (posterior and anterior division), temporal fusiform cortex (posterior division), temporal occipital fusiform cortex, planum polare, occipital fusiform gyrus, temporal pole, subcallosal cortex, and paracingulate gyrus. Morphological and functional changes in these areas have been previously reported in the literature. For example, the frontal, and temporal regions have been identified in many studies (Ibarretxe-Bilbao et al., 2012, Martin et al., 2009, Peng et al., 2017, Taylor et al., 1986, Tessa et al., 2014) and the insular cortex has been identified by Jubault et al. (2011). Overall, the regions identified by our model are in line with Braak’s staging model, which states that alphasynuclein deposition in the brain spreads in an ascending fashion, starting in the brainstem and extending towards the neocortex through the subcortex (subcortical regions) and mesocortex (parahippocampal cortex; (Braak et al., 2003)). Other cortical and subcortical regions were also identified in this work but with lower coverage and importance.

Many of the brain regions identified by the saliency map analysis were also selected by the RELIEFF feature selection approach when using Nu-SVM-RBF (e.g., parahippocampal gyrus, pallidum, putamen, temporal pole, caudate nucleus, insular cortex, subcallosal cortex, frontal orbital cortex, planum polare, occipital fusiform gyrus, paracingulate gyrus, and hippocampus). While this enhances confidence, it also suggests that CNNs can extract more complex and localized features from morphological data beyond simple regional volumes as used by the conventional machine learning models that likely results in the higher classification accuracy. This may, for example, be explained by CNN models being able to identify morphological changes that only affect certain areas of a brain region, which may be hidden in simple regional morphological parameters such as the volume.

4.3. Limitations

Some limitations of this work need to be discussed. First, although the use of multicenter data increases the robustness of the results and model generalizability, the retrospective nature of the data collection also means that there are differences regarding the specific inclusion/exclusion criteria, psychiatric co-morbidities, cognitive impairment (including dementia) as well as overall PD diagnosis criteria. This could have affected the representativeness of the collected PD population, particularly for earlier stages of the disease where diagnosis is more challenging. For this reason, the proposed model needs to be validated in more detail in prospective studies. Second, it may be useful to investigate other suitable deep learning architectures for this purpose. However, an extensive evaluation of a wide variety of CNN models was out of the scope of this work. Third, using multi-modal MRI datasets such as diffusion-tensor MRI or T2-weighted MRI may also improve the classification accuracy (Talai et al., 2021). Nevertheless, this work aimed to provide a simple single channel model focussed solely on T1-weighted datasets as this image sequence is widely available to enhance the feasibility of a clinical application in the future. Lastly, the subject’s T1-weighted MRI scans were registered to an atlas using the ANTs toolkit, which can be very time consuming. However, this process could also be performed with a deep learning approach like Voxelmorph (Balakrishnan et al., 2019) to speed-up the process, if needed. More importantly, there are other registration approaches besides ANTs that could have been implemented and would have consequently led to slightly different deformation fields due to differences in the algorithms and regularizer being used. This should be investigated in more detail in future research.

5. Conclusion

The present study demonstrates that 3D CNNs can be used to make an explainable computer-aided diagnosis of PD based on 3D Jacobian maps derived from morphological differences between groups. For this, a CNN was trained using preprocessed heterogeneous data from 12 separate imaging studies, increasing the models’ robustness. The model achieved a 79.3% accuracy, 80.2% precision, 81.3% specificity, 77.7% sensitivity, and 0.87 ROC-AUC on the holdout test data and showed generalizable performance on an independent test set. Moreover, the brain regions discovered are known to be associated with PD, providing reassurance for the results. The suggested model could function as a clinical decision support system, incorporating physician discretion and interpretation of the predictions (Maceachern and Forkert, 2021).

Funding

This work was supported by the Canadian Consortium on Neurodegeneration in Aging (CCNA), the Canada Research Chairs program, the River Fund at Calgary Foundation, and the Canadian Open Neuroscience Platform (CONP), the Canadian Institutes for Health Research, and the Tourmaline Chair in Parkinson disease.

CRediT authorship contribution statement

Milton Camacho: Conceptualization, Methodology, Software, Validation, Formal analysis, Investigation, Data curation, Writing – original draft, Writing – review & editing, Visualization. Matthias Wilms: Software, Writing – review & editing. Pauline Mouches: Software, Writing – review & editing. Hannes Almgren: Conceptualization, Writing – review & editing. Raissa Souza: Validation, Data curation. Richard Camicioli: Resources, Writing – review & editing. Zahinoor Ismail: Resources, Writing – review & editing. Oury Monchi: Conceptualization, Resources, Writing – review & editing. Nils D. Forkert: Conceptualization, Methodology, Resources, Writing – review & editing, Supervision.

Declaration of Competing Interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Acknowledgements

Image data used were provided in part by the OASIS-3 project (Prin-cipal Investigators: T. Benzinger, D. Marcus, J. Morris; NIH P50AG00561, P30 NS09857781, P01 AG026276, P01 AG003991, R01AG043434, UL1 TR000448, R01 EB009352), by the PPMI-a, public-private partnership funded by Michael J. Fox Foundation, by theOpenfMRI database (accession number ds000245), and by the UKBiobank (application number 77508). Additionally, Alexandru Hanganu, Iris Kathol, Myrlene Gee, Tracy Hammer, and Marisela Beas assisted in the conduction of this study through general advice, technical support, and proof reading.

Footnotes

2

Data used in the preparation of this article were obtained from the Parkinson’s Progression Markers Initiative (PPMI) database (www.ppmi-info.org/data). For up-to-date information on the study, visit www.ppmi-info.org

6

Pan-UKBB team. https://pan.ukbb.broadinstitute.org. 2020.

Appendix A

Supplementary data to this article can be found online at https://doi.org/10.1016/j.nicl.2023.103405.

Appendix A. Supplementary data

The following are the Supplementary data to this article:

Supplementary data 1
mmc1.docx (132.3KB, docx)

Data availability

Openly available datasets supporting the findings of this study include PPMI, C-BIG, OpenNeuro Japan, NEUROCON and Tao Wu data set. The other studies included retain ownership of their scans.

References

  1. Acharya H.J., Bouchard T.P., Emery D.J., Camicioli R.M. Axial signs and magnetic resonance imaging correlates in Parkinson’s disease. Can. J. Neurol. Sci. 2007;34:56–61. doi: 10.1017/S0317167100005795. [DOI] [PubMed] [Google Scholar]
  2. Adeli, E., Wu, G., Saghafi, B., An, L., Shi, F., Shen, D., 2017. Kernel-based joint feature selection and max-margin classification for early diagnosis of Parkinson’s disease. Sci. Reports 7: 1–14. 10.1038/srep41069. [DOI] [PMC free article] [PubMed]
  3. Adeli E., Shi F., An L., Wee C.Y., Wu G., Wang T., Shen D. Joint feature-sample selection and robust diagnosis of Parkinson’s disease from MRI data. Neuroimage. 2016;141:206–219. doi: 10.1016/J.NEUROIMAGE.2016.05.054. [DOI] [PMC free article] [PubMed] [Google Scholar]
  4. Aishwarya T., Ravi Kumar V. Machine learning and deep learning approaches to analyze and detect COVID-19: A review. SN Comput. Sci. 2021;2:1–9. doi: 10.1007/s42979-021-00605-9. [DOI] [PMC free article] [PubMed] [Google Scholar]
  5. Amoroso N., la Rocca M., Monaco A., Bellotti R., Tangaro S. Complex networks reveal early MRI markers of Parkinson’s disease. Med. Image Anal. 2018;48:12–24. doi: 10.1016/J.MEDIA.2018.05.004. [DOI] [PubMed] [Google Scholar]
  6. Avants B.B., Tustison N.J., Song G., Cook P.A., Klein A., Gee J.C. A reproducible evaluation of ANTs similarity metric performance in brain image registration. Neuroimage. 2011;54:2033. doi: 10.1016/J.NEUROIMAGE.2010.09.025. [DOI] [PMC free article] [PubMed] [Google Scholar]
  7. Badea L., Onu M., Wu T., Roceanu A., Bajenaru O. Exploring the reproducibility of functional connectivity alterations in Parkinson’s disease. PLoS One. 2017;12 doi: 10.1371/JOURNAL.PONE.0188196. [DOI] [PMC free article] [PubMed] [Google Scholar]
  8. Baecker L., Garcia-Dias R., Vieira S., Scarpazza C., Mechelli A. Machine learning for brain age prediction: Introduction to methods and clinical applications. EBioMedicine. 2021;72 doi: 10.1016/J.EBIOM.2021.103600. [DOI] [PMC free article] [PubMed] [Google Scholar]
  9. Balakrishnan G., Zhao A., Sabuncu M.R., Guttag J., Dalca A.V. VoxelMorph: A learning framework for deformable medical image registration. IEEE Trans. Med. Imaging. 2019;38:1788–1800. doi: 10.1109/TMI.2019.2897538. [DOI] [PubMed] [Google Scholar]
  10. Balestrino R., Schapira A.H.V. Parkinson disease. Eur. J. Neurol. 2020;27:27–42. doi: 10.1111/ENE.14108. [DOI] [PubMed] [Google Scholar]
  11. Boelmans K., Holst B., Hackius M., Finsterbusch J., Gerloff C., Fiehler J., Münchau A. Brain iron deposition fingerprints in Parkinson’s disease and progressive supranuclear palsy. Mov. Disord. 2012;27:421–427. doi: 10.1002/MDS.24926. [DOI] [PubMed] [Google Scholar]
  12. Braak H., del Tredici K., Rüb U., de Vos R.A.I., Jansen Steur E.N.H., Braak E. Staging of brain pathology related to sporadic Parkinson’s disease. Neurobiol. Aging. 2003;24:197–211. doi: 10.1016/S0197-4580(02)00065-9. [DOI] [PubMed] [Google Scholar]
  13. Chakraborty S., Aich S., Kim H.-C. Detection of Parkinson’s disease from 3T T1 weighted MRI scans using 3D convolutional neural network. Diagnostics. 2020;10 doi: 10.3390/DIAGNOSTICS10060402. [DOI] [PMC free article] [PubMed] [Google Scholar]
  14. Cigdem O., Yilmaz A., Beheshti I., Demirel H. 26th IEEE Signal Processing and Communications Applications Conference, SIU 2018. Institute of Electrical and Electronics Engineers Inc. 2018. Comparing the performances of PDF and PCA on Parkinson’s disease classification using structural MRI images; pp. 1–4. [DOI] [Google Scholar]
  15. Desikan R.S., Ségonne F., Fischl B., Quinn B.T., Dickerson B.C., Blacker D., Buckner R.L., Dale A.M., Maguire R.P., Hyman B.T., Albert M.S., Killiany R.J. An automated labeling system for subdividing the human cerebral cortex on MRI scans into gyral based regions of interest. Neuroimage. 2006;31:968–980. doi: 10.1016/J.NEUROIMAGE.2006.01.021. [DOI] [PubMed] [Google Scholar]
  16. Duchesne S., Chouinard I., Potvin O., Fonov V.S., Khademi A., Bartha R., Bellec P., Collins D.L., Descoteaux M., Hoge R., McCreary C.R., Ramirez J., Scott C.J.M., Smith E.E., Strother S.C., Black S.E. The Canadian Dementia Imaging Protocol: Harmonizing National Cohorts. J. Magn. Reson. Imaging. 2019;49:456–465. doi: 10.1002/JMRI.26197. [DOI] [PubMed] [Google Scholar]
  17. El-Manzalawy, Y., Honavar, V., 2005. WLSVM: integrating libsvm into weka environment. Software available at http://www. cs. iastate. edu/yasser/wlsvm.
  18. Esmaeilzadeh, S., Yang, Y., Adeli, E., 2018. End-to-end Parkinson disease diagnosis using brain MR-images by 3D-CNN.
  19. Fearnley J.M., Lees A.J. Ageing and Parkinson’s disease: Substantia nigra regional selectivity. Brain. 1991;114:2283–2301. doi: 10.1093/brain/114.5.2283. [DOI] [PubMed] [Google Scholar]
  20. Frazier J.A., Chiu S., Breeze J.L., Makris N., Lange N., Kennedy D.N., Herbert M.R., Bent E.K., Koneru V.K., Dieterich M.E., Hodge S.M., Rauch S.L., Grant P.E., Cohen B.M., Seidman L.J., Caviness V.S., Biederman J. Structural brain magnetic resonance imaging of limbic and thalamic volumes in pediatric bipolar disorder. Am. J. Psychiatry. 2005;162:1256–1265. doi: 10.1176/APPI.AJP.162.7.1256. [DOI] [PubMed] [Google Scholar]
  21. Goldstein J.M., Seidman L.J., Makris N., Ahern T., O’Brien L.M., Caviness V.S., Kennedy D.N., Faraone S.V., Tsuang M.T. Hypothalamic abnormalities in schizophrenia: sex effects and genetic vulnerability. Biol. Psychiatry. 2007;61:935–945. doi: 10.1016/J.BIOPSYCH.2006.06.027. [DOI] [PubMed] [Google Scholar]
  22. Hanganu A., Bedetti C., Degroot C., Mejia-Constain B., Lafontaine A.L., Soland V., Chouinard S., Bruneau M.A., Mellah S., Belleville S., Monchi O. Mild cognitive impairment is linked with faster rate of cortical thinning in patients with Parkinson’s disease longitudinally. Brain. 2014;137:1120–1129. doi: 10.1093/BRAIN/AWU036. [DOI] [PubMed] [Google Scholar]
  23. Ibarretxe-Bilbao N., Junque C., Segura B., Baggio H.C., Marti M.J., Valldeoriola F., Bargallo N., Tolosa E. Progression of cortical thinning in early Parkinson’s disease. Mov. Disord. 2012;27:1746–1753. doi: 10.1002/MDS.25240. [DOI] [PubMed] [Google Scholar]
  24. Isensee F., Schell M., Pflueger I., Brugnara G., Bonekamp D., Neuberger U., Wick A., Schlemmer H.P., Heiland S., Wick W., Bendszus M., Maier-Hein K.H., Kickingereder P. Automated brain extraction of multisequence MRI using artificial neural networks. Hum. Brain Mapp. 2019;40:4952. doi: 10.1002/HBM.24750. [DOI] [PMC free article] [PubMed] [Google Scholar]
  25. Jubault T., Gagnon J.F., Karama S., Ptito A., Lafontaine A.L., Evans A.C., Monchi O. Patterns of cortical thickness and surface area in early Parkinson’s disease. Neuroimage. 2011;55:462–467. doi: 10.1016/J.NEUROIMAGE.2010.12.043. [DOI] [PubMed] [Google Scholar]
  26. Kingma, D.P., Ba, J.L., 2017. Adam: A method for stochastic optimization. In: 3rd International Conference on Learning Representations, ICLR 2015 - Conference Track Proceedings. International Conference on Learning Representations, ICLR. 10.48550/arxiv.1412.6980.
  27. Kononenko, I., Šimec, E., Robnik-Šikonja, M., 1997. Overcoming the myopia of inductive learning algorithms with RELIEFF. Appl. Intell. 1997 7:1 7, 39–55. 10.1023/A:1008280620621.
  28. Laansma M.A., Bright J.K., Al-Bachari S., Anderson T.J., Ard T., Assogna F., Baquero K.A., Berendse H.W., Blair J., Cendes F., Dalrymple-Alford J.C., de Bie R.M.A., Debove I., Dirkx M.F., Druzgal J., Emsley H.C.A., Garraux G., Guimarães R.P., Gutman B.A., Helmich R.C., Klein J.C., Mackay C.E., McMillan C.T., Melzer T.R., Parkes L.M., Piras F., Pitcher T.L., Poston K.L., Rango M., Ribeiro L.F., Rocha C.S., Rummel C., Santos L.S.R., Schmidt R., Schwingenschuh P., Spalletta G., Squarcina L., van den Heuvel O.A., Vriend C., Wang J.J., Weintraub D., Wiest R., Yasuda C.L., Jahanshad N., Thompson P.M., van der Werf Y.D. International multicenter analysis of brain structure across clinical stages of Parkinson’s disease. Mov. Disord. 2021 doi: 10.1002/mds.28706. [DOI] [PMC free article] [PubMed] [Google Scholar]
  29. LaMontagne, P.J., Keefe, S., Lauren, W., Xiong, C., Grant, E.A., Moulder, K.L., Morris, J.C., Benzinger, T.L.S., Marcus, D.S., 2018. IC‐P‐164: OASIS‐3: longitudinal neuroimaging, clinical, and cognitive dataset for normal aging and Alzheimer disease. Alzheimers Dementia 14, 2019. 10.1016/j.jalz.2018.06.2231.
  30. Lang S., Hanganu A., Gan L.S., Kibreab M., Auclair-Ouellet N., Alrazi T., Ramezani M., Cheetham J., Hammer T., Kathol I., Sarna J., Monchi O. Network basis of the dysexecutive and posterior cortical cognitive profiles in Parkinson’s disease. Mov. Disord. 2019;34:893–902. doi: 10.1002/MDS.27674. [DOI] [PubMed] [Google Scholar]
  31. Leow A.D., Yanovsky I., Chiang M.-C., Lee A.D., Klunder A.D., Lu A., Becker J.T., Davis S.W., Toga A.W., Thompson P.M., Chiang C., Lee A.D., Klunder A.D., Lu A., Toga A.W., Thompson P.M., Becker J.T., Davis S.W. Statistical properties of Jacobian maps and the realization of unbiased large-deformation nonlinear image registration. IEEE Trans. Med. Imaging. 2007;26 doi: 10.1109/TMI.2007.892646. [DOI] [PubMed] [Google Scholar]
  32. lo Vercio, L., Amador, K., Bannister, J.J., Crites, S., Gutierrez, A., MacDonald, M.E., Moore, J., Mouches, P., Rajashekar, D., Schimert, S., Subbanna, N., Tuladhar, A., Wang, N., Wilms, M., Winder, A., Forkert, N.D., 2020. Supervised machine learning tools: A tutorial for clinicians. J Neural Eng 17, 062001. 10.1088/1741-2552/abbff2. [DOI] [PubMed]
  33. Maceachern S.J., Forkert N.D. Machine learning for precision medicine. Genome. 2021;64:416–425. doi: 10.1139/GEN-2020-0131/ASSET/IMAGES/LARGE/GEN-2020-0131F1.JPEG. [DOI] [PubMed] [Google Scholar]
  34. Makris N., Goldstein J.M., Kennedy D., Hodge S.M., Caviness V.S., Faraone S.V., Tsuang M.T., Seidman L.J. Decreased volume of left and total anterior insular lobule in schizophrenia. Schizophr. Res. 2006;83:155–171. doi: 10.1016/J.SCHRES.2005.11.020. [DOI] [PubMed] [Google Scholar]
  35. Martin W.R.W., Wieler M., Gee M., Camicioli R. Temporal lobe changes in early, untreated Parkinson’s disease. Mov. Disord. 2009;24:1949–1954. doi: 10.1002/MDS.22680. [DOI] [PubMed] [Google Scholar]
  36. Mouches P., Wilms M., Rajashekar D., Langner S., Forkert N.D. Multimodal biological brain age prediction using magnetic resonance imaging and angiography with the identification of predictive regions. Hum. Brain Mapp. 2022 doi: 10.1002/HBM.25805. [DOI] [PMC free article] [PubMed] [Google Scholar]
  37. Peng H., Gong W., Beckmann C.F., Vedaldi A., Smith S.M. Accurate brain age prediction with lightweight deep neural networks. Med. Image Anal. 2021;68 doi: 10.1016/J.MEDIA.2020.101871. [DOI] [PMC free article] [PubMed] [Google Scholar]
  38. Peng B., Wang S., Zhou Z., Liu Y., Tong B., Zhang T., Dai Y. A multilevel-ROI-features-based machine learning method for detection of morphometric biomarkers in Parkinson’s disease. Neurosci. Lett. 2017;651:88–94. doi: 10.1016/J.NEULET.2017.04.034. [DOI] [PubMed] [Google Scholar]
  39. Poewe, W., Seppi, K., Tanner, C.M., Halliday, G.M., Brundin, P., Volkmann, J., Schrag, A.E., Lang, A.E., 2017. Parkinson disease. Nature Rev. Dis. Primers 2017 3:1 3, 1–21. 10.1038/nrdp.2017.13. [DOI] [PubMed]
  40. Rizzo G., Copetti M., Arcuti S., Martino D., Fontana A., Logroscino G. Accuracy of clinical diagnosis of Parkinson disease. Neurology. 2016;86:566–576. doi: 10.1212/WNL.0000000000002350. [DOI] [PubMed] [Google Scholar]
  41. Sarasso E., Agosta F., Piramide N., Filippi M. Progression of grey and white matter brain damage in Parkinson’s disease: a critical review of structural MRI literature. J. Neurol. 2021;268:3144–3179. doi: 10.1007/S00415-020-09863-8/TABLES/4. [DOI] [PubMed] [Google Scholar]
  42. Sharma N., Sharma R., Jindal N. Machine Learning and Deep Learning Applications-A Vision. Glob. Trans. Proc. 2021;2:24–28. doi: 10.1016/j.gltp.2021.01.004. [DOI] [Google Scholar]
  43. Shinde, P.P., Shah, S., 2018. A review of machine learning and deep learning applications. In: Proceedings - 2018 4th International Conference on Computing, Communication Control and Automation, ICCUBEA 2018. Institute of Electrical and Electronics Engineers Inc. 10.1109/ICCUBEA.2018.8697857.
  44. Smilkov, D., Thorat, N., Kim, B., Viégas, F., Wattenberg, M., 2017. SmoothGrad: removing noise by adding noise. 10.48550/arxiv.1706.03825.
  45. Solana-Lavalle G., Rosas-Romero R. Classification of PPMI MRI scans with voxel-based morphometry and machine learning to assist in the diagnosis of Parkinson’s disease. Comput. Methods Programs Biomed. 2021;198 doi: 10.1016/J.CMPB.2020.105793. [DOI] [PubMed] [Google Scholar]
  46. Stanley E., Rajashekar D., Mouches P., Wilms M., Plettl K., Forkert N.D. A fully convolutional neural network for explainable classification of attention deficit hyperactivity disorder. SPIE-Intl. Soc. Opt. Eng. 2022:36. doi: 10.1117/12.2607509. [DOI] [Google Scholar]
  47. Talai A.S., Sedlacik J., Boelmans K., Forkert N.D. Utility of multi-modal MRI for differentiating of Parkinson’s disease and progressive supranuclear palsy using machine learning. Front. Neurol. 2021;12:546. doi: 10.3389/FNEUR.2021.648548/BIBTEX. [DOI] [PMC free article] [PubMed] [Google Scholar]
  48. Taylor A.E., Saint-cyr J.A., Lang A.E. Frontal lobe dysfunction in parkinson’s disease: The cortical focus of neostriatal outflow. Brain. 1986;109:845–883. doi: 10.1093/brain/109.5.845. [DOI] [PubMed] [Google Scholar]
  49. Tessa C., Lucetti C., Giannelli M., Diciotti S., Poletti M., Danti S., Baldacci F., Vignali C., Bonuccelli U., Mascalchi M., Toschi N. Progression of brain atrophy in the early stages of Parkinson’s disease: A longitudinal tensor-based morphometry study in de novo patients without cognitive impairment. Hum. Brain Mapp. 2014;35:3932–3944. doi: 10.1002/HBM.22449. [DOI] [PMC free article] [PubMed] [Google Scholar]
  50. Tibshirani R. Regression shrinkage and selection via the Lasso. J. R. Stat. Soc. Series B (Methodological) 1996;58:267–288. [Google Scholar]
  51. Tsiouris S., Bougias C., Konitsiotis S., Papadopoulos A., Fotopoulos A. Early-onset frontotemporal dementia-related semantic variant of primary progressive aphasia: multimodal evaluation with brain perfusion SPECT, SPECT/MRI coregistration, and MRI volumetry. Clin. Nucl. Med. 2022;47:260–264. doi: 10.1097/RLU.0000000000003934. [DOI] [PubMed] [Google Scholar]
  52. Tustison N.J., Avants B.B., Cook P.A., Zheng Y., Egan A., Yushkevich P.A., Gee J.C. N4ITK: improved N3 bias correction. IEEE Trans. Med. Imaging. 2010;29:1310–1320. doi: 10.1109/TMI.2010.2046908. [DOI] [PMC free article] [PubMed] [Google Scholar]
  53. Wei, D., Zhuang, K., Chen, Q., Yang, W., Liu, W., Wang, K., Sun, J., Qiu, J., 2017. Structural and functional MRI from a cross-sectional Southwest University Adult lifespan Dataset (SALD). bioRxiv. 10.1101/177279. [DOI] [PMC free article] [PubMed]
  54. Xiao, Y., Fonov, V., Chakravarty, M.M., Beriault, S., al Subaie, F., Sadikot, A., Pike, G.B., Bertrand, G., Collins, D.L., 2017. A dataset of multi-contrast population-averaged brain MRI atlases of a Parkinson׳s disease cohort. Data Brief 12, 370–379. 10.1016/j.dib.2017.04.013. [DOI] [PMC free article] [PubMed]

Associated Data

This section collects any data citations, data availability statements, or supplementary materials included in this article.

Supplementary Materials

Supplementary data 1
mmc1.docx (132.3KB, docx)

Data Availability Statement

Openly available datasets supporting the findings of this study include PPMI, C-BIG, OpenNeuro Japan, NEUROCON and Tao Wu data set. The other studies included retain ownership of their scans.

Articles from NeuroImage : Clinical are provided here courtesy of Elsevier

RESOURCES