3D Convolutional Network Research Articles

Brain MR images segmentation using 3D CNN with features recalibration mechanism for segmented CT generation

The segmentation of MR (magnetic resonance) images is a simple approach to create Pseudo CT images which are useful for many medical imaging analysis applications. One of the main challenges of this process is the bone segmentation of brain MR images. Deep convolutional neural networks (CNNs) have been widely and efficiently applied to perform MR images segmentation. The aim of this work is to propose a novel excitation-based CNN by recalibrating the network features adaptively to enhance the bone segmentation by segmenting the brain MR images into three tissue classes: bone, soft tissue, and air. The proposed method combines two types of features excitation mechanisms namely: (1) spatial squeeze and channel excitation block (cSE) and (2) channel squeeze and spatial excitation block (sSE). The two blocks are combined sequentially and integrated seamlessly into a 3D convolutional encoder decoder network. The novelty of this work emerges in the combination of the two excitation blocks sequentially to improve the segmentation performance and reduce the model complexity. The proposed approach is evaluated through a comparison with computed tomography (CT) images as ground truth and validated with other methods in the literature that applied deep CNN approaches to perform MR image segmentation for PET attenuation correction. Brain MR and CT datasets which consist of 50 patients are used to evaluate the proposed method. The segmentation performance of the three brain classes is evaluated using precision, recall, dice similarity coefficient (DSC), and Jaccard index. The presented method improves the bone tissue segmentation compared to the baseline model and other methods in the literature where the DSC is improved from 0.6278 ± 0.0006 to 0.6437 ± 0.0006 with an improvement percentage of 2.53% for bone class. The proposed excitation-based segmentation network architecture demonstrates promising and competitive results compared with other methods in the literature and reduces the model complexity thanks to the sequential combination of the two excitation blocks.

Micro-Attention Branch, a Flexible Plug-In That Enhances Existing 3D ConvNets

In the field of video motion recognition, with the increase in network depth, there is an asymmetry in the amount of parameters and their accuracy. We propose the structure of micro-attention branches and a method of integrating attention branches in multi-branches 3D convolution networks. Our proposed attention branches can improve this asymmetry problem. Attention branches can be flexibly added to 3D convolution network in the form of plug-in, without changing the overall structure of the original network. Through this structure, the newly constructed network can fuse the attention features extracted by attention branches in real time in the process of feature extraction. By adding attention branches, the model can focus on the action subject more accurately, so as to improve the accuracy of the model. Moreover, in the case that there are multiple sub-branches in the construction module of the existing network, 3D micro-attention branches can well adapt to this scenario. In the kinetics dataset, we use our proposed micro-attention branch structure to construct a deep network, which is compared with the original network. The experimental results show that the recognition accuracy of the network with micro-attention branches is improved by 3.6% compared with the original network, while the amount of parameters to be trained is only increased by 0.6%.

Multi-Scale Attention 3D Convolutional Network for Multimodal Gesture Recognition.

Gesture recognition is an important direction in computer vision research. Information from the hands is crucial in this task. However, current methods consistently achieve attention on hand regions based on estimated keypoints, which will significantly increase both time and complexity, and may lose position information of the hand due to wrong keypoint estimations. Moreover, for dynamic gesture recognition, it is not enough to consider only the attention in the spatial dimension. This paper proposes a multi-scale attention 3D convolutional network for gesture recognition, with a fusion of multimodal data. The proposed network achieves attention mechanisms both locally and globally. The local attention leverages the hand information extracted by the hand detector to focus on the hand region, and reduces the interference of gesture-irrelevant factors. Global attention is achieved in both the human-posture context and the channel context through a dual spatiotemporal attention module. Furthermore, to make full use of the differences between different modalities of data, we designed a multimodal fusion scheme to fuse the features of RGB and depth data. The proposed method is evaluated using the Chalearn LAP Isolated Gesture Dataset and the Briareo Dataset. Experiments on these two datasets prove the effectiveness of our network and show it outperforms many state-of-the-art methods.

Learning Unsupervised Visual Representations using 3D Convolutional Autoencoder with Temporal Contrastive Modeling for Video Retrieval

The rapid growth of tag-free user-generated videos (on the Internet), surgical recorded videos, and surveillance videos has necessitated the need for effective content-based video retrieval systems. Earlier methods for video representations are based on hand-crafted, which hardly performed well on the video retrieval tasks. Subsequently, deep learning methods have successfully demonstrated their effectiveness in both image and video-related tasks, but at the cost of creating massively labeled datasets. Thus, the economic solution is to use freely available unlabeled web videos for representation learning. In this regard, most of the recently developed methods are based on solving a single pretext task using 2D or 3D convolutional network. However, this paper designs and studies a 3D convolutional autoencoder (3D-CAE) for video representation learning (since it does not require labels). Further, this paper proposes a new unsupervised video feature learning method based on joint learning of past and future prediction using 3D-CAE with temporal contrastive learning. The experiments are conducted on UCF-101 and HMDB-51 datasets, where the proposed approach achieves better retrieval performance than state-of-the-art. In the ablation study, the action recognition task is performed by fine-tuning the unsupervised pre-trained model where it outperforms other methods, which further confirms the superiority of our method in learning underlying features. Such an unsupervised representation learning approach could also benefit the medical domain, where it is expensive to create large label datasets.

Individual Tree Detection in Urban ALS Point Clouds with 3D Convolutional Networks

Since trees are a vital part of urban green infrastructure, automatic mapping of individual urban trees is becoming increasingly important for city management and planning. Although deep-learning-based object detection networks are the state-of-the-art in computer vision, their adaptation to individual tree detection in urban areas has scarcely been studied. Some existing works have employed 2D object detection networks for this purpose. However, these have used three-dimensional information only in the form of projected feature maps. In contrast, we exploited the full 3D potential of airborne laser scanning (ALS) point clouds by using a 3D neural network for individual tree detection. Specifically, a sparse convolutional network was used for 3D feature extraction, feeding both semantic segmentation and circular object detection outputs, which were combined for further increased accuracy. We demonstrate the capability of our approach on an urban topographic ALS point cloud with 10,864 hand-labeled ground truth trees. Our method achieved an average precision of 83% regarding the common 0.5 intersection over union criterion. 85% percent of the stems were found correctly with a precision of 88%, while tree area was covered by the individual tree detections with an F1 accuracy of 92%. Thereby, we outperformed traditional delineation baselines and recent detection networks.

Subdivision-based Mesh Convolution Networks

Convolutionalneural networks (CNNs) have made great breakthroughs in two-dimensional (2D) computer vision. However, their irregular structure makes it hard to harness the potential of CNNs directly on meshes. A subdivision surface provides a hierarchical multi-resolution structure in which each face in a closed 2-manifold triangle mesh is exactly adjacent to three faces. Motivated by these two observations, this article presents SubdivNet , an innovative and versatile CNN framework for three-dimensional (3D) triangle meshes with Loop subdivision sequence connectivity. Making an analogy between mesh faces and pixels in a 2D image allows us to present a mesh convolution operator to aggregate local features from nearby faces. By exploiting face neighborhoods, this convolution can support standard 2D convolutional network concepts, e.g., variable kernel size, stride, and dilation. Based on the multi-resolution hierarchy, we make use of pooling layers that uniformly merge four faces into one and an upsampling method that splits one face into four. Thereby, many popular 2D CNN architectures can be easily adapted to process 3D meshes. Meshes with arbitrary connectivity can be remeshed to have Loop subdivision sequence connectivity via self-parameterization, making SubdivNet a general approach. Extensive evaluation and various applications demonstrate SubdivNet’s effectiveness and efficiency.

SPNet: A deep network for broadcast sports video highlight generation

Professionally broadcasted sports videos usually have long durations but contain only a few exciting events. In general, professional bodies and amateur content creators spend thousands of man-hours to manually crop the exciting video segments from these long-duration videos and generate handcrafted highlights. Sports enthusiasts keep them updated with the latest happening based on such highlights. There exists a need for a method that accurately and automatically recognizes the exciting activities in a sports game. To address this issue, we present a deep learning-based network SPNet that recognizes exciting sports activities by exploiting high-level visual feature sequences and automatically generates highlights. The proposed SPNet utilizes the strength of 3D convolution networks and Inception blocks for accurate activity recognition. We divide the sports video excitement into views, actions, and situations. Moreover, we provide 156 new annotations for about twenty-three thousand videos of the SP-2 dataset. Extensive experiments are conducted using two datasets SP-2 and C-sports, and the results demonstrate the superiority of the proposed SPNet. Our proposed method achieves the highest performance for views, action, and situation activities with an average accuracy of 76% on the SP-2 dataset and 82% on the C-sports dataset.

Attention-based 3D convolutional networks

ABSTRACT Being simple and portable, the three-dimensional (3D) convolution network has achieved great success in action recognition. However, its applicability in spatiotemporal feature learning is not evident. This study aims to improve the 3D convolution model and propose a flexible and significant attention module for the extraction of spatiotemporal information. Our first contribution is a self-additive attention module and a feature-based attention module, which is a simple yet effective method for measuring the spatiotemporal importance of a video. In self-additive attention, the spatiotemporal fusion between the frames is defined intuitively, where we set equivalent weights between the video frames manually. Further, the feature-based attention that is trained adaptively by the 3D convolution process combines the spatiotemporal information from the feature map. This study also focuses on attention fusion in learning the spatiotemporal characteristics for 3D convolution. The proposed attention fusion method exhibits outstanding performance in comparison to the recently developed attention modules and the latest 3D networks when applied to the data from the UCF101 and HMDB51 datasets. The experiments show consistent improvements, affirming the robustness of the method in extracting spatiotemporal attention.

Shape Carving Methods of Geologic Body Interpretation from Seismic Data Based on Deep Learning

The task of seismic data interpretation is a time-consuming and uncertain process. Machine learning tools can help to build a shortcut between raw seismic data and reservoir characteristics of interest. Recently, techniques involving convolutional neural networks have started to gain momentum. Convolutional neural networks are particularly efficient at pattern recognition within images, and this is why they are suitable for seismic facies classification and interpretation tasks. We experimented with three different architectures based on convolutional layers and compared them with different synthetic and field datasets in terms of quality of the seismic interpretation results and computational efficiency. The architectures used in our study were three deep fully convolutional architectures: a 3D convolutional network with a fully connected head; a 2D fully convolutional network, and U-Net. We found the U-Net architecture to be both robust and the fastest when performing classification at the prediction stage. The 3D convolutional model with a fully connected head was the slowest, while a fully convolutional model was unstable in its predictions.

Heart Rate Measurement Based on 3D Central Difference Convolution with Attention Mechanism.

Remote photoplethysmography (rPPG) is a video-based non-contact heart rate measurement technology. It is a fact that most existing rPPG methods fail to deal with the spatiotemporal features of the video, which is significant for the extraction of the rPPG signal. In this paper, we propose a 3D central difference convolutional network (CDCA-rPPGNet) to measure heart rate, with an attention mechanism to combine spatial and temporal features. First, we crop and stitch the region of interest together through facial landmarks. Next, the high-quality regions of interest are fed to CDCA-rPPGNet based on a central difference convolution, which can enhance the spatiotemporal representation and capture rich relevant time contexts by collecting time difference information. In addition, we integrate the attention module into the neural network, aiming to strengthen the ability of the neural network to extract video channels and spatial features, so as to obtain more accurate rPPG signals. In summary, the three main contributions of this paper are as follows: (1) the proposed network base on central difference convolution could better capture the subtle color changes to recover the rPPG signals; (2) the proposed ROI extraction method provides high-quality input to the network; (3) the attention module is used to strengthen the ability of the network to extract features. Extensive experiments are conducted on two public datasets—the PURE dataset and the UBFC-rPPG dataset. In terms of the experiment results, our proposed method achieves 0.46 MAE (bpm), 0.90 RMSE (bpm) and 0.99 R value of Pearson’s correlation coefficient on the PURE dataset, and 0.60 MAE (bpm), 1.38 RMSE (bpm) and 0.99 R value of Pearson’s correlation coefficient on the UBFC dataset, which proves the effectiveness of our proposed approach.

C3D-ConvLSTM based cow behaviour classification using video data for precision livestock farming

Cow behaviour provides valuable information about animal welfare, activities and livestock production. Therefore, monitoring of behaviour is gaining importance in the improvement of animal health, fertility and production yield. However, recognizing or classifying different behaviours with high accuracy is challenging, because of the high similarity of movements among these behaviours. In this study, we propose a deep learning framework to monitor and classify dairy behaviours, which is intelligently combined with C3D (Convolutional 3D) network and ConvLSTM (Convolutional Long Short-Term Memory) to classify the five common behaviours included feeding, exploring, grooming, walking and standing. For this, 3D CNN features were firstly extracted from video frames using C3D network; then ConvLSTM was applied to further extract spatio-temporal features, and the final obtained features were fed to a softmax layer for behaviour classification. The proposed approach using 30-frame video length achieved 90.32% and 86.67% classification accuracy on calf and cow datasets respectively, which outperformed the state-of-the-art methods including Inception-V3, SimpleRNN, LSTM, BiLSTM and C3D. Additionally, the influence of video length on behaviour classification was also investigated. It was found that increasing video sequence length to 30-frames enhanced classification performance. Extensive experiments show that combining C3D and ConvLSTM together can improve video-based behaviour classification accuracy noticeably using spatial–temporal features, which enables automated behaviour classification for precision livestock farming.

Learnable Empirical Mode Decomposition based on Mathematical Morphology

Empirical mode decomposition (EMD) is a fully data driven method for multiscale decomposing signals into a set of components known as intrinsic mode functions. EMD is based on lower and upper envelopes of the signal in an iterated decomposition scheme. In this paper, we put forward a simple yet effective method to learn EMD from data by means of morphological operators. We propose an end-to-end framework by incorporating morphological EMD operators into deeply learned representations, trained using standard backpropagation principle and gradient descent-based optimization algorithms. Three generalizations of morphological EMD are proposed: a) by varying the family of structuring functions, b) by varying the pair of morphological operators used to calculate the envelopes, and c) the use of a convex sum of envelopes instead of the classical mean point used in classical EMD. We discuss in particular the invariances that are induced by the morphological EMD representation. Experimental results on supervised classification of hyperspectral images by 1D convolutional networks demonstrate the interest of our method.

Transferable Features from 1D-Convolutional Network for Industrial Malware Classification

With the development of information technology, malware threats to the industrial system have become an emergent issue, since various industrial infrastructures have been deeply integrated into our modern works and lives. To identify and classify new malware variants, different types of deep learning models have been widely explored recently. Generally, sufficient data is usually required to achieve a well-trained deep learning classifier with satisfactory generalization ability. However, in current practical applications, an ample supply of data is absent in most specific industrial malware detection scenarios. Transfer learning as an effective approach can be used to alleviate the influence of the small sample size problem. In addition, it can also reuse the knowledge from pre-trained models, which is beneficial to the real-time requirement in industrial malware detection. In this paper, we investigate the transferable features learned by a 1D-convolutional network and evaluate our proposed methods on 6 transfer learning tasks. The experiment results show that 1D-convolutional architecture is effective to learn transferable features for malware classification, and indicate that transferring the first 2 layers of our proposed 1D-convolutional network is the most efficient way to reuse the learned features.

Improving Human Activity Recognition Integrating LSTM With Different Data Sources: Features, Object Detection and Skeleton Tracking

Over the past few years, technologies in the field of computer vision have greatly advanced. The use of deep neural networks, together with the development of computing capabilities, has made it possible to solve problems of great interest to society. In this work, we focus on one such problem that has seen a great development, the recognition of actions in live videos. Although the problem has been oriented in different ways in the literature, we have focused on indoor residential environments, such as a house or a nursing home. Our system can be used to understand what actions a person or group of people are carrying out. Two of the approaches used to solve the problem have been 3D convolution networks and recurrent networks. In our case, we have created a model that accurately combines several recurrent networks with processed data from different techniques: image feature extraction, object detection and people’s skeletons. The need to integrate these three techniques arises from the search to improve the detection of certain actions by taking advantage of the best recognition offered by each of the methods. In a complete experimentation, where several techniques have been evaluated against different datasets, the classification of the actions has been improved with respect to the existing models.

From Whole Video to Frames: Weakly-Supervised Domain Adaptive Continuous-Time QoE Evaluation.

Due to the rapid increase in video traffic and relatively limited delivery infrastructure, end users often experience dynamically varying quality over time when viewing streaming videos. The user quality-of-experience (QoE) must be continuously monitored to deliver an optimized service. However, modern approaches for continuous-time video QoE estimation require densely annotating the continuous-time QoE labels, which is labor-intensive and time-consuming. To cope with such limitations, we propose a novel weakly-supervised domain adaptation approach for continuous-time QoE evaluation, by making use of a small amount of continuously labeled data in the source domain and abundant weakly-labeled data (only containing the retrospective QoE labels) in the target domain. Specifically, given a pair of videos from source and target domains, effective spatiotemporal segment-level feature representation is first learned by a combination of 2D and 3D convolutional networks. Then, a multi-task prediction framework is developed to simultaneously achieve continuous-time and retrospective QoE predictions, where a quality attentive adaptation approach is investigated to effectively alleviate the domain discrepancy without hampering the prediction performance. This approach is enabled by explicitly attending to the video-level discrimination and segment-level transferability in terms of the domain discrepancy. Experiments on benchmark databases demonstrate that the proposed method significantly improves the prediction performance under the cross-domain setting.

Residue Assignment in Crystallographic Protein Electron Density Maps With 3D Convolutional Networks

This work proposes a neural network architecture, called 3D FC-DenseNet, for assigning amino acid labels to X-ray crystallographic electron density maps without relying on the amino acid sequence of proteins. The 3DFC-DenseNet is able to treat the task as a 3D semantic segmentation problem, assigning amino acid labels directly to protein electron density maps. By creating dedicated data sets and models for high, medium and low resolution samples, our method matches the performance of crystallographic toolkits for primary structure assignment at high resolutions. Furthermore, it outperforms them at medium resolution and functions at low resolutions where current toolkits and human ability fails.

Research on the Identification and Evaluation of Aerobics Movements Based on Deep Learning

A deep learning approach is used in this study to provide insight into aerobics movement recognition, and the model is used for aerobics movement recognition. The model complexity is significantly reduced, while the multi-scale features of the target at the fine-grained level are extracted, significantly improving the characterization of the target, by embedding lightweight multi-scale convolution modules in 3D convolutional residual networks to increase the local perceptual field range in each layer of the network. Finally, using the channel attention mechanism, the key features are extracted from the multi-scale features. To create a dual-speed frame rate detection model, the fast-slow combination idea is fused into a 3D convolutional network. To obtain spatial semantic information and motion information in the video, the model uses different frame rates, and the two-channel information is fused with features using lateral concatenation. Following the acquisition of all features, the features are fed into a temporal detection network to identify temporal actions and to design a behavior recognition system for the network model to demonstrate the network model's applicability. The average scores of students in the experimental group were significantly higher than those in the control group in seven areas: set accuracy, movement amplitude, movement strength, body coordination, coordination of movement and music, movement expression, and aesthetics; the average scores of movement proficiency and body control in the experimental group were also significantly higher than those in the control group, but the differences were not significant. The differences between the eight indicators in the experimental group were not significant when compared to those in the preexperimental group, indicating that intensive rhythm training for students improves secondary school students' comprehension, proficiency, and presentation of aerobics sets.

3D U-Net Improves Automatic Brain Extraction for Isotropic Rat Brain Magnetic Resonance Imaging Data.

Brain extraction is a critical pre-processing step in brain magnetic resonance imaging (MRI) analytical pipelines. In rodents, this is often achieved by manually editing brain masks slice-by-slice, a time-consuming task where workloads increase with higher spatial resolution datasets. We recently demonstrated successful automatic brain extraction via a deep-learning-based framework, U-Net, using 2D convolutions. However, such an approach cannot make use of the rich 3D spatial-context information from volumetric MRI data. In this study, we advanced our previously proposed U-Net architecture by replacing all 2D operations with their 3D counterparts and created a 3D U-Net framework. We trained and validated our model using a recently released CAMRI rat brain database acquired at isotropic spatial resolution, including T2-weighted turbo-spin-echo structural MRI and T2*-weighted echo-planar-imaging functional MRI. The performance of our 3D U-Net model was compared with existing rodent brain extraction tools, including Rapid Automatic Tissue Segmentation, Pulse-Coupled Neural Network, SHape descriptor selected External Regions after Morphologically filtering, and our previously proposed 2D U-Net model. 3D U-Net demonstrated superior performance in Dice, Jaccard, center-of-mass distance, Hausdorff distance, and sensitivity. Additionally, we demonstrated the reliability of 3D U-Net under various noise levels, evaluated the optimal training sample sizes, and disseminated all source codes publicly, with a hope that this approach will benefit rodent MRI research community.Significant Methodological Contribution: We proposed a deep-learning-based framework to automatically identify the rodent brain boundaries in MRI. With a fully 3D convolutional network model, 3D U-Net, our proposed method demonstrated improved performance compared to current automatic brain extraction methods, as shown in several qualitative metrics (Dice, Jaccard, PPV, SEN, and Hausdorff). We trust that this tool will avoid human bias and streamline pre-processing steps during 3D high resolution rodent brain MRI data analysis. The software developed herein has been disseminated freely to the community.

Neural marching cubes

We introduce Neural Marching Cubes , a data-driven approach for extracting a triangle mesh from a discretized implicit field. We base our meshing approach on Marching Cubes (MC), due to the simplicity of its input, namely a uniform grid of signed distances or occupancies, which frequently arise in surface reconstruction and from neural implicit models. However, classical MC is defined by coarse tessellation templates isolated to individual cubes. While more refined tessellations have been proposed by several MC variants, they all make heuristic assumptions, such as trilinearity, when determining the vertex positions and local mesh topologies in each cube. In principle, none of these approaches can reconstruct geometric features that reveal coherence or dependencies between nearby cubes (e.g., a sharp edge ), as such information is unaccounted for, resulting in poor estimates of the true underlying implicit field. To tackle these challenges, we re-cast MC from a deep learning perspective, by designing tessellation templates more apt at preserving geometric features, and learning the vertex positions and mesh topologies from training meshes, to account for contextual information from nearby cubes. We develop a compact per-cube parameterization to represent the output triangle mesh, while being compatible with neural processing, so that a simple 3D convolutional network can be employed for the training. We show that all topological cases in each cube that are applicable to our design can be easily derived using our representation, and the resulting tessellations can also be obtained naturally and efficiently by following a few design guidelines. In addition, our network learns local features with limited receptive fields, hence it generalizes well to new shapes and new datasets. We evaluate our neural MC approach by quantitative and qualitative comparisons to all well-known MC variants. In particular, we demonstrate the ability of our network to recover sharp features such as edges and corners, a long-standing issue of MC and its variants. Our network also reconstructs local mesh topologies more accurately than previous approaches. Code and data are available at https://github.com/czq142857/NMC.

Constrained harmonization algorithm for pooling multi‐site datasets

AbstractBackgroundPooling datasets from multiple studies can significantly improve statistical power: larger sample sizes can enable the identification of otherwise weak disease‐specific patterns. When modern learning methods are utilized (e.g., for predicting progression to dementia), differences in data acquisition‐methods / scanner‐protocols can enable the model to “cheat”, i.e. utilizes site‐specific artifacts rather than disease‐specific features. In this study, we develop a method to harmonize the performance of DNN classifiers across scanners/sites, via so‐called fairness constraints, thereby encouraging consistent behavior while controlling for site‐specific nuisance variables.MethodWe conducted two studies: (a) to demonstrate feasibility of pooling across sites (Site‐Pooling) and (b) to pool data across scanners (Scanner‐Pooling). For Site‐Pooling, our analysis included summaries from Freesurfer processed T1‐weighted images of the Wisconsin Alzheimer's Disease Research Center (ADRC) and German Center for Neurodegenerative Diseases (DZNE). The Freesurfer summaries were used to train a two layer neural network classifier and five‐fold cross‐validation performance was assessed. For Scanner‐Pooling experiments, Freesurfer processed MR images from the Alzheimer's Disease Neuroimaging Initiative (ADNI) were used to train a deep 3D convolutional network. Performance average on a held‐out test dataset was evaluated. In both cases, a constraint to equalize the performance of the trained classifier across the domains (sites/scanners) was incorporated during training.ResultTable 1 shows the results of AD/MCI classification for site‐pooling analysis. Our proposed method is compared against a naive pooling approach which does not incorporate the “harmonization constraint”. As shown, the proposed method improves the “difference of errors” measure by 8% / 7% and with only a small drop in overall error rates. Figure 1 illustrates the results from our scanner‐pooling analysis. The performance across the three scanners, GE, Siemens and Philips, is evaluated pair‐wise. A consistent improvement in harmonization is observed and only ∼2% drop in overall error rate is seen.ConclusionWe provide a harmonization constraint based algorithm to mitigate site specific differences when performing analysis of pooled brain imaging datasets in AD studies. In contrast to a method which modifies the data, we achieve harmonization by constraining the classifier to perform similarly across sites/groups/scanners, improving reproducibility.