Differential Diagnosis of Frontotemporal Dementia and Alzheimer's Disease using Generative Adversarial Network

summary

The gist

Differential Diagnosis of Frontotemporal Dementia and Alzheimer’s Disease using Generative Adversarial Network is a study proposing a novel framework combining multiscale structural features from…

In short

This study uses a Generative Adversarial Network (GAN) combined with multiscale structural features from MRI scans to differentiate between Frontotemporal Dementia (FTD), Alzheimer’s Disease (AD), and normal controls. The framework achieved 87.80% accuracy, demonstrating that combining multi-scale data analysis and synthetic data augmentation significantly improves the ability to distinguish between these dementia subtypes.

Key concepts

Multiscale Structural Features
The researchers analyzed brain MRI scans by dividing them into patches of different sizes (500, 1000, and 2000 voxels). Extracting features at these different scales helps capture distinguishing information that might be present in both very small and very large structural changes within the brain tissue.
Generative Adversarial Network (GAN)
A GAN consists of two parts: a generator that creates synthetic data and a discriminator that tries to tell real data from fake data. In this study, the generator was trained on MRI patches to create realistic samples for AD, FTD, and normal controls. This synthetic data helps improve the model's performance when the original MRI dataset is limited.
Multiscale Deep Neural Network (MDNN)
The MDNN acts as a sophisticated discriminator in the GAN setup. It is designed to learn complex latent patterns from the features extracted at various structural scales. By using this network, the system can better understand and classify brain structures based on these multi-scale representations, leading to more accurate multi-class differentiation.
Ensemble Classification Framework
To ensure robust results and prevent overfitting with limited data, the study trained ten separate classification networks. The final diagnosis is determined by taking the majority vote among these ten models. This ensemble approach provides a more stable and reliable classification performance for distinguishing between AD, FTD, and normal controls.

Terminology used across episodes

This episode discusses

The paper

Differential Diagnosis of Frontotemporal Dementia and Alzheimer's Disease using Generative Adversarial Network · Read on arXiv

Simon Fraser University · Tencent

Frontotemporal dementia and Alzheimer's disease are two common forms of dementia and are easily misdiagnosed as each other due to their similar pattern of clinical symptoms. Differentiating between the two dementia types is crucial for determining disease-specific intervention and treatment. Recent development of Deep-learning-based approaches in the field of medical image computing are delivering some of the best performance for many binary classification tasks, although its application in differential diagnosis, such as neuroimage-based differentiation for multiple types of dementia, has not been explored. In this study, a novel framework was proposed by using the Generative Adversarial Network technique to distinguish FTD, AD and normal control subjects, using volumetric features extracted at coarse-to-fine structural scales from Magnetic Resonance Imaging scans. Experiments of 10-folds cross-validation on 1,954 images achieved high accuracy. With the proposed framework, we have demonstrated that the combination of multi-scale structural features and synthetic data augmentation based on generative adversarial network can improve the performance of challenging tasks such as differentiating Dementia sub-types.

Transcript

Introduction to the show: ident: Electrical Engineering. The day in electrical engineering: signals, systems and the hardware that carries them.

Rosa: I'm Rosa.

Dev: And I'm Dev, here with Taro.

Rosa: We're looking at the paper: "Differential Diagnosis of Frontotemporal Dementia and Alzheimer's Disease using Generative Adversarial Network".

Dev: Here's the gist of the paper. Differential Diagnosis of Frontotemporal Dementia and Alzheimer’s Disease using Generative Adversarial Network is a study proposing a novel framework combining multiscale structural features from MRI scans with Generative Adversarial Networks to distinguish between Frontotemporal Dementia (FTD), Alzheimer’s disease (AD), and normal controls, which is crucial for determining disease-specific interventions <ref:2109.05627#pg2> The proposed framework achieved 87.80% accuracy through 10 fold cross validation experiment <ref:2109.05627#pg4> Frontotemporal dementia (FTD) and Alzheimer’s disease (AD) share similar clinical symptoms, leading to common misdiagnoses between the two forms <ref:2109.05627#pg2> The specificity in the distinguishing of AD from FTD is still low even with significant expertise from clinicians <ref:2109.05627#pg2> Volumetric structural change measured by T1-weighted MRI is an effective biomarker for differential diagnosis <ref:2109.05627#pg3> Most previous dementia classification studies focused on binary classification, such as normal control (NC) vs.

Rosa: FTD or NC vs.

Dev: AD <ref:2109.05627#pg3> To the best of our knowledge, no deep learning based approach has been applied to the differential diagnosis of AD and FTD <ref:2109.05627#pg4> Data Preparation and Feature Extraction A total of 1954 Structural MRI were used for the preparation of this article <ref:2109.05627#pg4> Only images from subject with stable diagnosis during the longitudinal follow up were included to eliminate potential uncertainty <ref:2109.05627#pg4> Image preprocessing involved several steps to reduce the dimension of input data by extracting preliminary features <ref:2109.05627#pg4> 1) We first segmented the gray matter of each T1 structural MRI image into 87 anatomical structures (FreeSurfer 5.3

4: ) <ref:2109.05627#pg4> 2) Each structures were regarded as a regions of interest (ROI) and were further parcellated into smaller subregions, which we termed as "patches" <ref:2109.05627#pg4> 3) Volume size and cortical thickness for each patch was extracted as features for downstream classification <ref:2109.05627#pg4> The features were extracted at three different scales, i.e., 500, 1000 and 2000 voxels per patch because the discriminate information among each class could exist in different scale of images The resulted subparcellation of the structural ROIs resulted a total of 1488, 705 and 343 patches across the entire brain, respectively <ref:2109.05627#pg4> Model Architecture and Training A Generative Adversarial Network (GAN) was trained with the patch-wise features extracted at multiscales from MRI to augment data samples <ref:2109.05627#pg5> The GAN consisted of a generator (G) and a discriminator (D), where G aimed to find a mapping from random variables to the distribution of the actual data, while D aimed to distinguish the candidates synthesized by the generator, denoted here as "fake", from the samples belong to true classes <ref:2109.05627#pg5> A Multiscale Deep Neural Network (MDNN) was applied as the discriminator to learn latent pattern from the feature space and perform multi-class classification <ref:2109.05627#pg5> The MDNN consisted of a cascade of four four-layer Multilayer Perceptrons (MLPs) <ref:2109.05627#pg5> Experimental Validation and Results Ten fold cross validation was performed on the 1954 T1 MRI images to validate the discriminant ability of the proposed framework <ref:2109.05627#pg4> Five deep learning based experiments were performed, including MLP with single scale feature extracted at 500, 1000, and 2000 voxels per patch, MDNN with multiscale features, and GAN with multiscale features The overall classification accuracy of the GAN approach was 87.80% The network performances well for the task of distinguishing between AD and FTD, while the discrimination between NC and FTD showed least accurate performance The cross validation experiments proved that the MDNN could learn the latent representation among features extracted at different structural scales from MRI scans <ref:2109.05627#pg5> Further improvement on discriminant ability can be achieved by leveraging the GAN for data augmentation <ref:2109.05627#pg5> The combination of multiscale structural features and synthetic data augmentation based on generative adversarial network can improve the performance of challenging tasks such as differentiating Dementia sub-types <ref:2109.05627#pg2> The final classification result of testing set were obtained through majority voting of the 10 networks References Alladi, S., Xuereb, J., Bak, T., Nestor, P., Knibb, J., Patterson, K., Hodges, J.: Focal cortical presentations of Alzheimer’s disease <ref:2109.05627#pg2> Arjovsky, M., Bottou, L.: Towards principled methods for training generative adversarial networks <ref:2109.05627#pg5> Beg, M.F., Miller, M.I., Trouvé, A., Younes, L.: Computing large deformation metric mappings via geodesic flows of diffeomorphisms <ref:2109.05627#pg5> Dale AM, Fischl B, S.M.: Cortical surface-based analysis <ref:2109.05627#pg5> Davatzikos, C., Resnick, S.M., Wu, X., Parmpi, P., Clark, C.M.: Individual patient diagnosis of AD and FTD via high-dimensional pattern classification of MRI <ref:2109.05627#pg5> Du, A.T., Schuff, N., Kramer, J.H., Rosen, H.J., Gorno-Tempini, M.L., Rankin, K., Miller, B.L., Weiner, M.W.: Different regional patterns of cortical thinning in Alzheimer’s disease and frontotemporal dementia <ref:2109.05627#pg5> Neary, D., Snowden, J., Mann, D.: Frontotemporal dementia <ref:2109.05627#pg2> Neary, D., Snowden, J.S., Gustafson, L., Passant, U., Stuss, D., Black, S., Freedman, M., Kertesz, A., Robert, P., Albert, M.: Frontotemporal lobar degeneration A consensus on clinical diagnostic criteria <ref:2109.05627#pg2> Pasquier, F.: Telling the difference between frontotemporal dementia and Alzheimer’s disease <ref:2109.05627#pg2> Raamana, P.R., Rosen, H., Miller, B., Weiner, M.W., Wang, L., Beg, M.F.: Threeclass differential diagnosis among Alzheimer disease, frontotemporal dementia, and controls <ref:2109.05627#pg2> Raamana, P.R., Weiner, M.W., Wang, L., Beg, M.F., Initiative, A.D.N., et al.: Thickness network features for prognostic applications in dementia <ref:2109.05627#pg5> Rabinovici, G., Seeley, W., Kim, E., Gorno-Tempini, M., Rascovsky, K., Pagliaro, T., Allison, S., Halabi, C., Kramer, J., Johnson, J.: Distinct MRI atrophy patterns in autopsy-proven Alzheimer’s disease and frontotemporal lobar degeneration <ref:2109.05627#pg5> Lu, D., Popuri, K., Ding, G.W., Balachandar, R., Beg, M.F.: Multiscale deep neural network based analysis of FDG-PET images for the early diagnosis of Alzheimer’s disease <ref:2109.05627#pg5> Ma, D., Popuri, K., Bhalla, M., Sangha, O., Lu, D., Cao, J., Jacova, C., Wang, L., Beg, M.F.: Quantitative assessment of field strength <ref:2109.05627#pg5> Neary, D., Snowden, J.S.: Evaluation of the NINCDS-ADRDA criteria in the differentiation of Alzheimer’s disease and frontotemporal dementia <ref:2109.05627#pg2> Page 1 --- Differential Diagnosis of Frontotemporal Dementia and Alzheimer’s Disease using Generative Adversarial Network Da Ma, Donghuan Lu, Karteek Popuri, and Mirza Faisal Beg Simon Fraser University, School of Engineering Science, Burnaby, V5A 1S6, Canada Tencent, Jarvis Lab <ref:2109.05627#pg2> Abstract.

Rosa: Frontotemporal dementia and Alzheimer’s disease are two common forms of dementia and are easily misdiagnosed as each other due to their similar pattern of clinical symptoms <ref:2109.05627#pg2> Differentiating between the two dementia types is crucial for determining disease-specific intervention and treatment <ref:2109.05627#pg2> Recent development of Deep-learning-based approaches in the field of medical image computing are delivering some of the best performance for many binary classification tasks, although its application in differential diagnosis, such as neuroimage-based differentiation for multiple types of dementia, has not been explored <ref:2109.05627#pg2> In this study, a novel framework was proposed by using the Generative Adversarial Network technique to distinguish FTD, AD and normal control subjects, using volumetric features extracted at coarse-to-fine structural scales from Magnetic Resonance Imaging scans <ref:2109.05627#pg2> Experiments of 10-folds cross validation on 1,954 images achieved high accuracy <ref:2109.05627#pg4> With the proposed framework, we have demonstrated that the combination of multiscale structural features and synthetic data augmentation based on generative adversarial network can improve the performance of challenging tasks such as differentiating Dementia sub-types <ref:2109.05627#pg2> Keywords: Differential Diagnosis, Magnetic Resonance Imaging, Generative Adversarial Network <ref:2109.05627#pg2> Page 3 --- Structural magnetic Resonance Imaging (MRI) is commonly used to detect pathological atrophic changes in the brain for diagnosis of neurodegenerative diseases such as AD and FTD <ref:2109.05627#pg4> The atrophy patterns of these two types of dementia differs in terms of affected regions and rate of change, and the volumetric structural change measured by T1-weighted MRI is an effective biomarker for differential diagnosis

5: <ref:2109.05627#pg4> Example of volumetric-based features includes gray matter (GM) volume loss

16: , cortical thinning

6: , High-dimensional features based on GM and white matter (WM) volumes distribution of whole brain

5: , as well as atrophy and shape deformity of individual structures

8: <ref:2109.05627#pg4> Most previous dementia classification studies focused on binary classification, e.g., normal control (NC) vs.

Title and authors: Rosa: So what exactly is the main goal here, beyond just trying to separate the diagnoses? What’s the actual mechanism they’re proposing?

Dev: They are using a novel framework that combines multiscale structural information from MRI scans with a Generative Adversarial Network to distinguish FTD, AD, and normal control subjects.

Rosa: It sounds complicated because of those "multiscale structural features" they mention. Can you explain that simply for someone who isn't an image processing expert?

Dev: They take the T1-weighted MRI images and first segment them into eighty-seven anatomical structures, then they further break those structures down into smaller patches. They extract features like volume size and cortical thickness for each patch at three different sizes: five hundred, one thousand and two thousand voxels per patch.

Rosa: So they’re looking at the brain from a big view down to really fine details in those patches?

Dev: Precisely. They say that discriminating information among each class could exist at different scales of images, so they are testing features across these different levels. They end up with a total of one thousand four hundred eighty-eight seven hundred five, and three hundred forty-three patches across the whole brain depending on the scale used.

Taro: From my side, I’m wondering about how much noise those smaller patch sizes introduce when you're trying to find a pattern in real tissue atrophy.

Rosa: That’s a good point Taro. But they argue that this multiscale approach is necessary because the distinguishing information might be hidden at different structural levels in the scans.

The paper's summary: Dev: So, what do we actually get from putting all those features—the multiscale patches—into a model? What’s the overall finding they report on this study?

Rosa: The main result they report is that when they use this proposed framework, combining the multiscale structural features and synthetic data augmentation based on the Generative Adversarial Network, they achieved an accuracy of eighty-seven point eight zero percent through ten fold cross validation.

Dev: That eighty-seven point eight zero percent is a solid number for a multi-class classification task involving these complex brain diseases. It shows that this combination of methods works quite well for the task of differentiating between AD and FTD.

Rosa: The paper specifically points out that while the GAN helps improve performance, there’s a specific nuance in how it performs the discrimination between normal controls and Frontotemporal Dementia subjects; that distinction showed less accurate performance.

Dev: That tells us something important about where the current framework is strongest and weakest, even with all those advanced techniques applied. They also note that the Multiscale Deep Neural Network was successful in learning a latent pattern from the feature space using those features extracted at different scales.

Taro: It seems like they’ve built a robust system for general differentiation, but maybe the subtlety between the most similar group, AD and FTD, is still a challenge for this specific setup.

The paper's improvements: Rosa: So we’re talking about what makes this framework better than just using a single scale of features or just a standard deep learning model? What are the key improvements they suggest?

Dev: The paper suggests two main things that help push the performance up. First, using the multiscale features helps keep as much information as possible during feature dimension reduction, which lets them capture discriminating information across different structural levels in the scans.

Rosa: And second, they use the Generative Adversarial Network for data augmentation. They train a generator and a discriminator to synthesize new MRI samples based on the patch-wise features extracted at multiscale.

Dev: That augmentation is meant to address the limited number of available MRI scans in this study, which is a real constraint when you’re training deep learning models for rare conditions. They found that this combination of multiscale structural features and synthetic data augmentation can improve performance on challenging tasks like differentiating dementia subtypes by about twenty percent.

Taro: So, the augmentation isn't just padding the dataset; it's actually helping the discriminator learn a better representation of what true data looks like across those various scales. That sounds more useful than just adding random noise.

Rosa: It really does sound like they’re trying to make the model more resilient to variations in how these diseases manifest structurally, which is important when you’re moving from a lab setting to real-world trials.

Conclusion: Dev: So, wrapping this up, what’s the big picture for clinicians or researchers looking at this paper on "Differential Diagnosis of Frontotemporal Dementia and Alzheimer's Disease using Generative Adversarial Network"?

Rosa: Basically, they show that combining multiscale structural features with synthetic data augmentation via a GAN can improve the performance of challenging tasks like differentiating dementia subtypes. It suggests this integrated approach is useful for moving beyond the binary classification studies that have dominated the field so far.

Dev: They proved that this combination helps learn latent representations across different structural scales from MRI scans, and they used an ensemble classification framework by training ten networks with independently separated splits to get a final result through majority voting.

Taro: I just want to say that while eighty-seven point eight zero percent is good, the paper also flags a limitation: the discrimination between normal controls and Frontotemporal Dementia showed least accurate performance. That means we still need better tools for those specific comparisons.

Rosa: True, so this framework is strong for AD versus FTD, but they admit it’s not perfect when trying to pin down the normal control group specifically. It opens the door for future work focusing on improving that specific separation.

Dev: Exactly. So we’ve seen how they used multiscale features and GAN augmentation to boost accuracy, and while it’s a solid result for dementia sub-type differentiation, the next step is definitely refining those comparisons where it falls short.

More episodes

← Home