3DSiameseNet to Analyze Brain MRI
Prediction of the cognitive evolution of a person susceptible to develop a neurodegenerative disorder is crucial to provide an appropriate treatment as soon as possible. In this paper we propose a 3D siamese network designed to extract features from whole-brain 3D MRI images. We show that it is possible to extract meaningful features using convolution layers, reducing the need of classical image processing operations such as segmentation or pre-computing features such as cortical thickness. To lead this study we used the Alzheimer’s Disease Neuroimaging Initiative (ADNI), a public data base of 3D MRI brain images. A set of 247 subjects has been extracted, all of the subjects having 2 images in a range of 12 months. In order to measure the evolution of the patients states we have compared these 2 images. Our work has been inspired at the beginning by an article of Bhagwat et al. in 2018, who have proposed a siamese network to predict the status of patients but without any convolutional layers and reducing the MRI images to a vector of features extracted from predefined ROIs. We show that our network achieves an accuracy of 90% in the classification of cognitively declining VS stable patients. This result has been obtained without the help of a cognitive score and with a small number of patients comparing to the current datasets size claimed in deep learning domain.
Cecilia Ostertag1-2, Marie
Beurton-Aimar2, Thierry Urruty3
\authoraddrfor the Alzheimer’s Disease Neuroimaging Initiative*
1 L3i - EA 2118 LaRochelle University, France, email@example.com
2 LaBRI - CNRS 5800 Bordeaux University, France, firstname.lastname@example.org
3 Univ. Poitiers, CNRS, XLIM, UMR 7252, F-86000 Poitiers, France, email@example.com
*Data used in preparation of this article were obtained from the Alzheimer’s Disease Neuroimaging Initiative (ADNI) database (adni.loni.usc.edu). As such, the investigators within the ADNI contributed to the design and implementation of ADNI and/or provided data but did not participate in analysis or writing of this report. A complete listing of ADNI investigators can be found at: http://adni.loni.usc.edu/wp- content/uploads/how_to_apply/ADNI_Acknowledgement_List.pdf
deep learning, siamese network, brain MRI, neurodegenerative disease, convolutional neural network
Neurodegenerative diseases like Alzheimer, Parkinson, …are disorders resulting in brain tissue damage. Because of the heavy impact on the brain morphology in this kind of diseases, medical imaging techniques such as structural MRI can be used to study them, and to distinguish between healthy patients, patients affected by disease or patients with mild cognitive impairment. This way, a computational model fitted to this task can use an MRI image to give a diagnosis. Beyond the diagnosis, for a better care of affected people, it becomes necessary to predict the evolution of the disease, to adapt the treatment of already identified patients, and to identify which patients are at high risk and have to be watched closely[frisoni2010clinical].
To predict evolution of such diseases in subjects with regular yearly medical examinations, we have designed a 3D siamese neural network taking into account the MRI images of two time points to predict their cognitive evolution. Our model is inspired by the work of Bhagwat et. al. [bhagwat2018modeling] who have proposed a siamese network to predict the evolution of patients by using measures computed from MRI images and a cognitive score. On one hand, as we would like to take into account 3D MRI, our network is made of several layers of convolutional layers for each of its branches. . On the other hand, our task has been to find if there are differences between two images of the same patients (at two different times), so the layer merging the two siamese branches has been defined to extract differences and not similarities as is usually the case with siamese models. In the next sections, we will first detail the architecture of our network, then we will present our results, and finally we will give insights to the discriminative abilities of our model.
2 Related works
The success of Deep Neural Networks (DNNs) for image classification tasks was shown by LeCun in 1998 [lecun1998gradient]. DNNs are multi-layer architectures designed to extract meaningful features in data, in order to solve a given task. These models are trained and optimized using gradient backpropagation: a loss function is computed to quantify the error of the model at each iteration, and the training phase aims to minimize this error by adjusting the model’s parameters in order to enhance its performance step by step. During training, the loss function is computed on a training dataset and on an independent validation dataset. The difference is that the network’s weights are not updated when using validation data, so the validation loss evaluates the performance of the network on unknown data. This method is used to ensure that the model is not overfitting the data.
To process image data, variations of DNNs, called Convolutional Neural Networks (CNNs) are most often used. They consist of a sequence of convolution layers, followed by pooling layers. The successive convolution operations select the most important features of the images, thus replacing the computation of hand-crafted features, and the pooling operations reduce the dimensions of the data. After the convolution layers, a classifier made of fully connected (or dense) layers, use these informations to predict classes.
To face the challenge of image pair comparison, another type of DNNs are convolutive siamese networks [koch2015siamese]. They are made of two identical branches, that share the same weights at initialization and during all training. A distance or similarity function is then used to compare two inputs, and based on this the network will be trained to distinguish similar pairs from dissimilar pairs. Such networks have been used for example for face recognition [zagoruyko2015learning, lin2016homemade].
MRI data are medical images representing in 3D the anatomy of an organ or tissue. The standard DNN architectures used with this type of volumetric data are CNNs. In the case of MRI brain images, for example Alzheimer’s Disease data, several input formats have been used to train these networks: 2D slices from the MRI volumes [sarraf2016classification, billones2016demnet], whole 3D volumes [hosseini2016alzheimer, payan2015predicting, korolev2017residual, khvostikov20183d], or regions of interest (ROIs) selected using an atlas [khvostikov20183d]. Also, a 2D+ approach used the three anatomical plane projections was proposed [aderghal2017classification]. Using 3D ROIs is effective to reduce memory requirements, but implies a priori knowledge of the brain areas involved in the disease. The model that we present in this work does not rely on ROI selection.
In all previously cited articles, the classification between the different groups of patients is inferred from only one time point (one MRI image), so the diagnostic that is predicted is only valid for that time point. To predict the symptomatic evolution of Alzheimer’s Disease in subjects, Bhagwat et. al. [bhagwat2018modeling] present a deep learning architecture which uses MRI and clinical data from two medical visits. To handle the two MRI images, they use a siamese network and create a ”distance embedding” to model the changes between the two visits. The main drawback of this model is that it is dedicated to use 1D vectors of measures computed from MRI data instead of directly using MRI images. According to the work of Payan and Montana [payan2015predicting] a 3D approach has better performances than a 2D approach because it is able to extract 3D patterns. The model that we propose is a siamese network as suggested by Bhagwat et. al. but with some major differences to take into account input pairs of 3D images.
Our architecture is based on the siamese network part of Bhagwat et. al. Logintudinal Siamese Network (LSN) [bhagwat2018modeling]. In this article the goal of the authors is to predict the evolution of Alzheimer’s disease using two classes, ”Stable” and ”Decline”. Obviously, ”Stable” represents the subject with no cognitive decline, and ”Decline” the subjects with cognitive decline. To create these classes, the authors used a clustering technique to separate the data based on a cognitive score. The LSN network is composed of two parallel branches, made of fully connected layers, and take as input a pair of 1D vectors. The first vector contains 78 cortical thickness values computed from the baseline visit MRI of a subject, and the second vector contains 78 cortical thickness values computed from the 12-month follow-up visit of the same subject. To obtain these values, the MRI images are segmented in 78 ROIs, and the cortical thickness value of each ROI is estimated. At the end of both branches, the two outputs are concatenated into one. Contrarily to usual siamese networks, this model is not used to evaluate the similarity between the two inputs to determine if they are from the same person, but to evaluate the difference between the two inputs to quantify the changes in the brain morphology induced by the disease. As stated previously we want to use directly 3D MRI images, so we have adapted the LSN model to process 3D inputs.
3 Network model
Image processing operations as segmentation and computation of cortical thickness steps are time consuming, often require arbitrary selection of ROIs, and can be subject to errors. To avoid this, we decided to adapt the LSN architecture [bhagwat2018modeling] to whole MRI 3D inputs. As we faced problems of memory and computation time due to the use of large and tridimentional data, we had several choices to make during the network implementation. Moreover, the choice of layers and parameter values were also conditioned by the need to extract small local features that are characteristic of changes in the brain.
The LSN model uses 1D inputs, so their siamese network is only made of fully-connected layers. To use 3D data, we replaced the fully-connected layers by 3D convolution layers, followed by 3D pooling layers. For all convolution layers we chose a kernel size of 3x3x3 and a stride of 1 to preserve the details of the images. We used average pooling instead of max pooling to avoid a bias of the network towards edge detection. In Fig.1 we detail the elementary block of our network.
It is made of one convolution layer, one batch normalization layer [ioffe2015batch], one leaky ReLU [maas2013rectifier], and one average pooling layer. This block is repeated three times in each branch of our network.
Because of memory allocation problems, we used a small number of filters for the convolution layers. The fusion of the two branches of the network was done by computing the difference instead of using concatenation, to have a better representation of the evolution of the subject’s brain morphology. Finally we used three fully-connected layers, with softmax activation for the last layer. The complete architecture of our 3D-SiameseNet is presented in Fig.2. To avoid overfitting a dropout of 50 % is applied before the output layer. Each layer is followed by batch normalization and leaky ReLU to improve the convergence of our model, as suggested by Lin et. al. [lin2016homemade].
3.2 Preprocessing and data augmentation
The data we used come from Alzheimer’s Disease Neuroimaging Initiative (ADNI) database [noauthor_adni_nodate]. We used skull-stripping [skull] to remove features non related to brain morphology, and zero-padding to ensure that all images have the same dimensions. The pixel values were scaled between 0 and 1, and the images were downscaled by 2 using an order 3 spline interpolation to reduce memory allocation. The final dimensions of our inputs are 102x108x75. For training we used pairs of images: the baseline MRI and the 12-month follow-up MRI. (see Fig.3)
Our dataset contains 247 image pairs. We take 40 images for validation and the rest in training. Because of the modest size of our training dataset compared to deep learning requirements, we used a data augmentation strategy. It consisted of random left/right rotation between 0°and 5°, and left/right axial symmetry. During training, data augmentation is applied image per image, at every epoch, on the training images only.
3.3 Training protocol
To assign labels to our data, we followed the script provided by Bhagwat et. al. [bhagwat2018modeling], subjects from the ADNI1 cohort were used for trajectory modelling, dividing the dataset between ”Stable” and ”Decline” subjects, based on the Mini Mental State Evaluation clinical scores. We selected a subset of these subjects in our dataset: 134 ”Stable” subjects, and 113 ”Decline” subjects. For all of these subjects we have a baseline and a 12-month MRI.
The model was trained with categorical crossentropy loss, optimized with Adam [kingma2014adam], with a base learning rate of 0.001. To reduce overfitting we applied L2 regularization to all kernel weights. We used 10 times random sub-sampling validation to assess the bias induced by the choice of validation data. The model was trained for 800 epochs, with the training data shuffled after every epoch, and using a batch size of 20 due to memory limitations. To evaluate the model, we used accuracy and Mean Squared Logarithmic Error (MSLE).
The model was implemented using Keras [chollet2015keras] with TensorFlow [tensorflow2015-whitepaper] backend, and run on two NVIDIA 2080 Ti GPUs. The source code can be found at https://github.com/morphoboid/3D-SiameseNet
Our network’s loss reaches a plateau after about 600 epochs (see Fig.4), with a training loss of 0.624 and a validation loss of 0.992 at the end of training.
We used accuracy to evaluate our model’s ability to predict the right class, and MSLE to evaluate the probabilities associated to these predictions. After training, our model achieves a mean accuracy of 90% and a mean MSLE of 0.04 (See Table 1).
Bhagwat et. al. did performance evaluation experiments using 1116 subjects, to compare their results with four other models: a logistic regression (LR), a support vector machine (SVM), a random forest (RF), and an artificial neural network (ANN). These four models only use imaging data (baseline MRI + follow-up MRI). Their LSM uses imaging and clinical data from two time points, and achieved 94% accuracy (see Table 2). While we cannot at this time compare our results to the LSM network because we only use imaging data, we show that our model outperforms LR, SVM, RF, and ANN models, using significantly less images.
Along with a high accuracy, the MSLE over the validation set is low, showing that the probabilities associated to the predictions are close to the true labels. This should ensure a small number of false positives and false negatives during inference.
As we mentioned in Section 3, we have been mainly limited by the memory size of our GPU clusters. With more memory-expansive computing resources, we could use MRI inputs at their original dimensions and avoid downscaling artifacts. We could also add filters to the convolution layers to extract more features. With this configuration, we assume that our model’s accuracy would further increase.
5 Discriminative abilities of our model
Inspired by the work of Hosseini-Asl et. al. [hosseini2016alzheimer], and to study in more depth the discriminative abilities of our model, we plotted the t-distributed Stochastic Neighbor Embedding (t-SNE) projection [maaten2008visualizing], a nonlinear dimensionality reduction algorithm, of the feature maps of 40 validation images obtained at three stages of our network (see layers circled in red in Fig.5). The feature maps we chose are the concatenated outputs of both left and right branches, the output of the Subtract layer, and the output of the second Dense layer (see Fig.2). We also plotted the concatenation of the input image pairs for comparison. In these projections, similar feature maps are drawn closer, and dissimilar feature maps are farther apart.
Fig.5 resumes the computing steps from the input step, square number to the final result, square number . We can observe at the beginning, the formation of three clusters, square number , that are not probably related to our two classes ”Stable” (Negative), and ”Decline” (Positive). We can hypothesize that they are due to other important characteristics of the MRI images, that were extracted by the network, as the brain size or another anatomical characteristics. Given that Bhagwat et. al. used a concatenation layer to fuse the two branches, there could be a bias introduced to their network at this step.
After the next layer, these three clusters are not identifiable anymore (see Fig.5.3), which could mean that the Subtract layer had the effect of removing the unwanted features of our pair of inputs.
Finally, at the penultimate Dense layer, the feature maps are grouped in two clusters, corresponding to our two classes (see Fig.5.3). We showed false positives and false negatives in additional colors to show how this embedding explains the classification error of the model. Comparing this result to the plot of the input pair, we can clearly see the discriminative ability of our network.
In this paper we have proposed 3D-SiameseNet, a deep learning model to analyze 3D brain MRI images in order to compare evolution between two time points. The main contribution of this model is the use of raw 3D MRI data as input, suppressing the need for beforehand segmentation and computation of manually selected features. Our architecture was designed taking into account memory constraints as well as the need to select local and precise features. By repeating the training and validation phase 10 times, we obtain a mean validation accuracy of 90% after training.
In future works we would like to add clinical information to our network to enhance its performance. The sub-network handling clinical data will have to be designed to extract features from the raw measures, before fusion with the imaging features.
This project received the financial support from ICE project (RCB 2018-A02699-46)
Data used in preparation of this article were obtained from the Alzheimer’s Disease
Neuroimaging Initiative (ADNI) database (adni.loni.usc.edu). As such, the investigators
within the ADNI contributed to the design and implementation of ADNI and/or provided
data but did not participate in analysis or writing of this report. A complete listing of
ADNI investigators can be found at:
Data used in the preparation of this article were obtained
from the Alzheimer’s Disease Neuroimaging Initiative (ADNI)
database. The primary goal of ADNI has been to test whether serial
magnetic resonance imaging (MRI), positron emission tomography
(PET), other biological markers, and clinical and
neuropsychological assessment can be combined to measure the
progression of mild cognitive impairment (MCI) and early Alzheimer’s
disease (AD).For up-to-date information, see www.adni-info.org.
Data collection and sharing for this project was funded by the Alzheimer’s Disease Neuroimaging Initiative (ADNI) (National Institutes of Health Grant U01 AG024904) and DOD ADNI (Department of Defense award number W81XWH-12-2-0012). ADNI is funded by the National Institute on Aging, the National Institute of Biomedical Imaging and Bioengineering, and through generous contributions from the following: AbbVie, Alzheimer’s Association; Alzheimer’s Drug Discovery Foundation; Araclon Biotech; BioClinica, Inc.; Biogen; Bristol-Myers Squibb Company; CereSpir, Inc.; Cogstate; Eisai Inc.; Elan Pharmaceuticals, Inc.; Eli Lilly and Company; EuroImmun; F. Hoffmann-La Roche Ltd and its affiliated company Genentech, Inc.; Fujirebio; GE Healthcare; IXICO Ltd.; Janssen Alzheimer Immunotherapy Research & Development, LLC.; Johnson & Johnson Pharmaceutical Research & Development LLC.; Lumosity; Lundbeck; Merck & Co., Inc.; Meso Scale Diagnostics, LLC.; NeuroRx Research; Neurotrack Technologies; Novartis Pharmaceuticals Corporation; Pfizer Inc.; Piramal Imaging; Servier; Takeda Pharmaceutical Company; and Transition Therapeutics. The Canadian Institutes of Health Research is providing funds to support ADNI clinical sites in Canada. Private sector contributions are facilitated by the Foundation for the National Institutes of Health (www.fnih.org). The grantee organization is the Northern California Institute for Research and Education, and the study is coordinated by the Alzheimer’s Therapeutic Research Institute at the University of Southern California. ADNI data are disseminated by the Laboratory for Neuro Imaging at the University of Southern California.