Hongyi Chen1, Xueling Liu2, Yuxin Li1,2, Puyeh Wu3, and Daoying Geng1,2
1Academy for Engineering and Technology, Fudan University, Shanghai, China, 2Huashan Hospital, State Key Laboratory of Medical Neurobiology, Fudan University, Shanghai, China, 3GE Healthcare, Beijing, China
Synopsis
Keywords: Data Analysis, Parkinson's Disease
In this study, we constructed a hybrid machine
learning model utilizing CNN and radiomics features based on NM-sensitive
setMag images. The hybrid features improved the diagnostic performance in
distinguishing PD patients from HC, as demonstrated in the SVM classifier,
which demonstrated 95.7% accuracy, 92.9% sensitivity, and 100% specificity. The
interpretability of the radiomics approach is better because radiomics features
provide more interpretable biomarkers, while the CNN approach extracts deeper
features from images. Furthermore, visualizing regions that influence classification
decisions via saliency map can also enhance the interpretability of the CNN
approach.
Background
Progressive loss of neuromelanin (NM) containing
dopamine-producing neurons and increase in iron deposition in substantia
nigra pars compacta (SNpc) are reliable and promising biomarkers for the
diagnosis of Parkinson's disease (PD)1,2. Due
to high iron and low NM content in SNpc, the diagnosis of PD patients can be
made quickly by calculating the reflected signal intensity and contrast ratio
(CR) on NM-sensitive MRI images3-5. However,
these conventional methods rely on manual delineation of the ROI6,7,
which is subjective, labor-intensive and error-prone. Recently, Takahashi et
al.8
proposed an automatic voxel-based SNpc segmentation system, while it was
subject to errors in the alignment process and insensitive to small SNpc
regions.
With
the rapid development of deep learning and convolutional neural network (CNN), it
has become a powerful tool for medical image applications9. However,
there are several obstacles to the application of CNN in PD diagnosis. Due to
the black-box nature of CNN models, this predisposes the diagnoses made by the
models to be less interpretable. Secondly, CNNs that utilize image features
cannot be effectively combined with histological features. Accordingly, researchers have started to
focus on the combination of CNNs and radiomics.
Choi et al.10
developed a fully automated model for segmentation and status prediction of
gliomas by utilizing tumor MRI images and radiomics features, and achieved an satisfactory prediction accuracy
in the external test.
In this study, NM-sensitive
short-echo-time magnitude (setMag) MRI images11 were firstly reconstructed as inputs. We proposed a hybrid model with machine
learning
classifier, utilizing features extracted from CNN model and radiomics approach, and aimed to verify
its feasibility and performance in PD diagnosis.Methods
This study was approved by the Ethics
Committee of our hospital (No. KY 2016–214). Informed consent was obtained from
all participants. 73 PD patients and 48 healthy controls (HC) were
recruited in this study. MRI examinations were performed on a 3.0-T
MR scanner (MR750; GE Healthcare, Milwaukee,
WI). 3D multi-echo GRE images were acquired, and NM-sensitive setMag images
were reconstructed following the procedure published previously11. Bilateral
hyperintensity ROIs in SNpc were manually depicted on setMag images using
ITK-SNAP12.
For the radiomics
model, we extracted 107 radiomics features from original images using PyRadiomics,
consisting of 6 categories: (1) first-order statistics; (2) GLCM; (3) GLDM; (4)
GLRLM; (5) GLSZM; (6) NGTDM. Each original image was subjected to 7
transformations including: wavelet, LoG, square, square-root, logarithm, exponential,
and gradient. In total, 1781 radiomics features were taken into consideration,
and were further selected by LASSO to retain the most important features.
For the CNN model, YOLO v5 model was used to
extract brainstem regions, and our CNN
model was modified from the LeNet-513. Specifically,
input images were sequentially passed through two sets of convolutional and
maximum pooling layers, before being spanned into a vector in the final
fully-connected layer.
Figure 1 shows entire process of the fusion and classification. In order to filter
CNN features, we quantified the importance of features using the MDI evaluation
method based on the random forest classifier. Top 20 features were selected and
connected to the radiomics features. Subsequently, features were divided into
training and validation sets at a ratio of 8:2. 7 algorithms were adopted
including: k-nearest neighbor, support vector machine, random forests, logistic
regression, Gaussian naïve Bayes, adaptive boosting and multilayer perceptron.
Depending on the performance, multiple index system and classifier models were
analyzed to determine the optimal model.Results
The image set extracted by YOLO v5 model was fed
into the CNN model for extracting features. After screening and fusion, we
obtained a total of 121 cases with a feature vector of size 1 × 54 for each
case. The results of the hybrid machine learning model on the validation set were
shown in Table 1. In addition, we plotted ROC curves and calculated the
AUC for each machine learning method (Figure 2). The best performing classifier
of hybrid model was SVM: ACC = 0.957, SEN = 0.929, SPE = 1.000, PPV = 1.000,
NPV = 0.900, F1-score = 0.963, and AUC = 0.99 (95% CI: 0.97–1.00), which was
improved comparing to the radiomics model: ACC = 0.913, SEN = 0.929,
SPE = 0.889, PPV = 0.929, NPV = 0.889, F1-score = 0.929, AUC = 0.94 (95% CI:
0.86–1.00).
To explain the role of CNN features, we
visualized the features extracted in the fully-connected layer of the CNN
model. As shown in Figure 3, the focus of the saliency map was on the
case characteristic "dovetail" of PD patients, which was consistent
with the pathological change regions of PD patients in the current study and
greatly increased the interpretability of our CNN model.Conclusion
In this study, we proposed a hybrid machine
learning model utilizing CNN and radiomics features for PD diagnosis. Our
results demonstrated that this approach produced the desired outcome. Radiomics
feature provided abundant texture information from images, and features
extracted by CNN were continuously optimized based on the classification
accuracy on the training set, enabling better problem solutions from
data-driven perspective. Furthermore, according to the saliency map, different locations
displayed varying degrees of significance which proves the interpretability of
our hybrid classification model.Acknowledgements
No acknowledgement found.References
- Ward, R. J., Zucca, F. A., Duyn, J. H., Crichton, R. R. & Zecca,
L. The role of iron in brain ageing and neurodegenerative disorders. The Lancet Neurology 13, 1045-1060 (2014).
- He, N. et al. Imaging iron and neuromelanin
simultaneously using a single 3D gradient echo magnetization transfer sequence:
Combining neuromelanin, iron and the nigrosome-1 sign as complementary imaging
biomarkers in early stage Parkinson's disease. Neuroimage 230, 117810,
doi:10.1016/j.neuroimage.2021.117810 (2021).
- Gaurav, R. et al. Longitudinal Changes in
Neuromelanin MRI Signal in Parkinson's Disease: A Progression Marker. Mov Disord 36, 1592-1602, doi:10.1002/mds.28531 (2021).
- Fabbri, M. et al. Substantia Nigra Neuromelanin as
an Imaging Biomarker of Disease Progression in Parkinson's Disease. J Parkinsons Dis 7, 491-501, doi:10.3233/JPD-171135 (2017).
- Prasad, S. et al. Three-dimensional
neuromelanin-sensitive magnetic resonance imaging of the substantia nigra in
Parkinson's disease. Eur J Neurol 25, 680-686, doi:10.1111/ene.13573
(2018).
- Taniguchi, D. et al. Neuromelanin imaging and
midbrain volumetry in progressive supranuclear palsy and Parkinson's
disease. 33, 1488-1492 (2018).
- Pyatigorskaya,
N. et al. Magnetic Resonance Imaging
Biomarkers to Assess Substantia Nigra Damage in Idiopathic Rapid Eye Movement
Sleep Behavior Disorder. Sleep 40, doi:10.1093/sleep/zsx149 (2017).
- Takahashi, H. et al. Comprehensive MRI quantification
of the substantia nigra pars compacta in Parkinson's disease. Eur J Radiol 109, 48-56, doi:10.1016/j.ejrad.2018.06.024 (2018).
- Le Berre, A. et al. Convolutional neural
network-based segmentation can help in assessing the substantia nigra in
neuromelanin MRI. Neuroradiology 61, 1387-1395,
doi:10.1007/s00234-019-02279-w (2019).
- Choi, Y. S. et al. Fully automated hybrid approach
to predict the IDH mutation status of gliomas via deep learning and radiomics. Neuro-oncology 23, 304-313 (2021).
- Liu, X. et al. Short-echo-time magnitude image
derived from quantitative susceptibility mapping could resemble
neuromelanin-sensitive MRI image in substantia nigra. 20,
262, doi:10.1186/s12883-020-01828-8 (2020).
- Yushkevich, P. A. et al. User-guided 3D active contour
segmentation of anatomical structures: significantly improved efficiency and
reliability. Neuroimage 31, 1116-1128 (2006).
- LeCun, Y. LeNet-5,
convolutional neural networks. URL: http://yann. lecun. com/exdb/lenet 20, 14 (2015).