Unsupervised Domain Adaptation Techniques Research Articles

A New Method for Bearing Fault Diagnosis across Machines Based on Envelope Spectrum and Conditional Metric Learning.

In recent years, most research on bearing fault diagnosis has assumed that the source domain and target domain data come from the same machine. The differences in equipment lead to a decrease in diagnostic accuracy. To address this issue, unsupervised domain adaptation techniques have been introduced. However, most cross-device fault diagnosis models overlook the discriminative information under the marginal distribution, which restricts the performance of the models. In this paper, we propose a bearing fault diagnosis method based on envelope spectrum and conditional metric learning. First, envelope spectral analysis is used to extract frequency domain features. Then, to fully utilize the discriminative information from the label distribution, we construct a deep Siamese convolutional neural network based on conditional metric learning to eliminate the data distribution differences and extract common features from the source and target domain data. Finally, dynamic weighting factors are employed to improve the convergence performance of the model and optimize the training process. Experimental analysis is conducted on 12 cross-device tasks and compared with other relevant methods. The results show that the proposed method achieves the best performance on all three evaluation metrics.

Triple Feature Disentanglement for One-Stage Adaptive Object Detection

In recent advancements concerning Domain Adaptive Object Detection (DAOD), unsupervised domain adaptation techniques have proven instrumental. These methods enable enhanced detection capabilities within unlabeled target domains by mitigating distribution differences between source and target domains. A subset of DAOD methods employs disentangled learning to segregate Domain-Specific Representations (DSR) and Domain-Invariant Representations (DIR), with ultimate predictions relying on the latter. Current practices in disentanglement, however, often lead to DIR containing residual domain-specific information. To address this, we introduce the Multi-level Disentanglement Module (MDM) that progressively disentangles DIR, enhancing comprehensive disentanglement. Additionally, our proposed Cyclic Disentanglement Module (CDM) facilitates DSR separation. To refine the process further, we employ the Categorical Features Disentanglement Module (CFDM) to isolate DIR and DSR, coupled with category alignment across scales for improved source-target domain alignment. Given its practical suitability, our model is constructed upon the foundational framework of the Single Shot MultiBox Detector (SSD), which is a one-stage object detection approach. Experimental validation highlights the effectiveness of our method, demonstrating its state-of-the-art performance across three benchmark datasets.

Local–global pseudo-label correction for source-free domain adaptive medical image segmentation

Domain shift is a commonly encountered issue in medical imaging solutions, primarily caused by variations in imaging devices and data sources. To mitigate this problem, unsupervised domain adaptation techniques have been employed. However, concerns regarding patient privacy and potential degradation of image quality have led to an increased focus on source-free domain adaptation. In this study, we address the issue of false labels in self-training based source-free domain adaptive medical image segmentation methods. To correct erroneous pseudo-labels, we propose a novel approach called the local–global pseudo-label correction (LGDA) method for source-free domain adaptive medical image segmentation. Our method consists of two components: An offline local context-based pseudo-label correction method that utilizes local context similarity in image space. And an online global pseudo-label correction method based on class prototypes, which corrects erroneously predicted pseudo-labels by considering the relative distance between pixel-wise feature vectors and prototype vectors. We evaluate the performance of our method on three benchmark fundus image datasets for optic disc and cup segmentation. Our method achieves superior performance compared to the state-of-the-art approaches, even without using any source data.

Domain adaptive remote sensing image semantic segmentation with prototype guidance

Current unsupervised domain adaptation (UDA) techniques in semantic segmentation effectively decrease the domain discrepancy between the labeled source domain and unlabeled target domain, thereby enhancing the model’s pixel-wise discriminative capability for target domain data. However, in remote sensing images (RSIs), our study uncovers that these approaches perform poorly in the presence of class distribution inconsistencies between the source and target domains. In this work, we propose a one-stage mean teacher framework with a novel auxiliary prototype classifier, named MTA, to address this issue. Specifically, the teacher model assigns pseudo labels at pixel level for target samples and captures knowledge from the student model via exponential moving average (EMA). With labeled source samples and target samples that have pseudo labels, the student model can alleviate the divergence between the source and target domains. In addition, the auxiliary prototype classifier (APC) reduces the performance degradation in the parametric softmax classifier of the student model caused by class distribution divergence. We also propose a prototype computation scheme to obtain each class prototype in the APC. Specifically, we build a memory bank for each class of the two domains to store feature embeddings dynamically. Then, we compute the class prototype by applying the clustering algorithm on memory banks corresponding to the class. Meanwhile, the APC reduces the intra-class domain discrepancy by optimizing the cross-entropy loss, which brings each class feature distribution of the two domains closer to the class prototype. The experimental results on RSIs UDA semantic segmentation tasks show the superiority of our approach over comparative methods.

Monitoring of Laser Powder Bed Fusion process by bridging dissimilar process maps using deep learning-based domain adaptation on acoustic emissions

Advances in sensorization and identification of information embedded inside sensor signatures during manufacturing processes using Machine Learning (ML) algorithms for better decision-making have become critical enablers in building data-driven monitoring systems. In the Laser Powder Bed Fusion (LPBF) process, data-driven-based process monitoring is gaining popularity since it allows for real-time component quality verification. Real-time qualification of the additively manufactured parts has a significant advantage as the cost of conventional post-manufacturing inspection methods can be reduced. Also, corrective actions or build termination could be done to save machine time and resources. However, despite the successful development in addressing monitoring needs in LPBF processes, less research has been paid to the ML model's robustness in decision-making when dealing with variations in data distribution from the laser-material interaction owing to different process spaces. Inspired by the idea of domain adaptation in ML, in this work, we propose a deep learning-based unsupervised domain adaptation technique to tackle shifts in data distribution owing to different process parameter spaces. The temporal waveforms of acoustic emissions from the LPBF process zone corresponding to three regimes, namely Lack of Fusion, conduction, and keyhole, were acquired on two different 316 L stainless steel powder distributions (> 45 µm and < 45 µm) with two different parameter sets. Temporal and spectral analysis of the acoustic waveforms corresponding to the powder distributions treated with different laser parameters showed the presence of offset in the data distribution, which was subsequently treated with the proposed unsupervised domain adaptation technique to have an ML model that could be generalized. Furthermore, the prediction accuracy of the proposed methodology between the two distributions showed the feasibility of adapting to the newer environment unsupervisedly and improving the ML model's generalizability.

Unsupervised Domain Adaptation Fundus Image Segmentation via Multi-scale Adaptive Adversarial Learning.

Segmentation of the Optic Disc (OD) and Optic Cup (OC) is crucial for the early detection and treatment of glaucoma. Despite the strides made in deep neural networks, incorporating trained segmentation models for clinical application remains challenging due to domain shifts arising from disparities in fundus images across different healthcare institutions. To tackle this challenge, this study introduces an innovative unsupervised domain adaptation technique called Multi-scale Adaptive Adversarial Learning (MAAL), which consists of three key components. The Multi-scale Wasserstein Patch Discriminator (MWPD) module is designed to extract domain-specific features at multiple scales, enhancing domain classification performance and offering valuable guidance for the segmentation network. To further enhance model generalizability and explore domain-invariant features, we introduce the Adaptive Weighted Domain Constraint (AWDC) module. During training, this module dynamically assigns varying weights to different scales, allowing the model to adaptively focus on informative features. Furthermore, the Pixel-level Feature Enhancement (PFE) module enhances low-level features extracted at shallow network layers by incorporating refined high-level features. This integration ensures the preservation of domain-invariant information, effectively addressing domain variation and mitigating the loss of global features. Two publicly accessible fundus image databases are employed to demonstrate the effectiveness of our MAAL method in mitigating model degradation and improving segmentation performance. The achieved results outperform current state-of-the-art (SOTA) methods in both OD and OC segmentation. Codes are available at https://github.com/M4cheal/MAAL.

Compositional Semantic Mix for Domain Adaptation in Point Cloud Segmentation.

Deep-learning models for 3D point cloud semantic segmentation exhibit limited generalization capabilities when trained and tested on data captured with different sensors or in varying environments due to domain shift. Domain adaptation methods can be employed to mitigate this domain shift, for instance, by simulating sensor noise, developing domain-agnostic generators, or training point cloud completion networks. Often, these methods are tailored for range view maps or necessitate multi-modal input. In contrast, domain adaptation in the image domain can be executed through sample mixing, which emphasizes input data manipulation rather than employing distinct adaptation modules. In this study, we introduce compositional semantic mixing for point cloud domain adaptation, representing the first unsupervised domain adaptation technique for point cloud segmentation based on semantic and geometric sample mixing. We present a two-branch symmetric network architecture capable of concurrently processing point clouds from a source domain (e.g.synthetic) and point clouds from a target domain (e.g.real-world). Each branch operates within one domain by integrating selected data fragments from the other domain and utilizing semantic information derived from source labels and target (pseudo) labels. Additionally, our method can leverage a limited number of human point-level annotations (semi-supervised) to further enhance performance. We assess our approach in both synthetic-to-real and real-to-real scenarios using LiDAR datasets and demonstrate that it significantly outperforms state-of-the-art methods in both unsupervised and semi-supervised settings.

Source-free Style-diversity Adversarial Domain Adaptation with Privacy-preservation for person re-identification

Unsupervised domain adaptation (UDA) techniques for person re-identification (ReID) have been extensively studied to facilitate the transfer of knowledge from labeled source domains to unlabeled target domains. However, the need to access the source data raises privacy concerns in real-world scenarios. To overcome this limitations, source-free domain adaptation (SFDA) was introduced, enabling adaptation without requiring access to the source data, relying instead on a well-trained source model. Nevertheless, existing SFDA methods assume a shared label space and overlook the significance of domain-style discrepancies in person ReID, limiting their applicability to source-free domain adaptive person ReID. In this paper, we present a novel approach called Source-free Style-diversity Adversarial Domain Adaptation with Privacy-preservation (S2ADAP) for person ReID to address these challenges. Our approach effectively handles inter-domain pedestrian appearance style differences using GAN-based domain-style diversity augmentation and intra-domain individual style misalignment through adversarial mutual teaching learning, avoiding access to data from the source domain. We leverage a pre-trained model as a person appearance style encoder to enhance source-similar style diversity in the target domain and achieve intra-domain individual style alignment by introducing the domain style discriminator to promote the discriminability of person semantic features for domain adaptation. The experimental results on publicly available person ReID datasets affirm the efficacy of our approach, offering a promising and privacy-preserving solution for person ReID tasks.

Single-source UDA for privacy-preserving intelligent fault diagnosis based on domain augmentation

In practical applications of fault diagnosis, several factors, including fluctuations in load, changes in equipment condition, and environmental noise effects, could cause a classifier that’s been trained on the source domain to be ill-suited for matching data from the target domain. Unsupervised domain adaptation techniques have been developed to tackle this issue, but they typically demand access to fully labeled source domains, ignoring concerns of privacy regarding source domain data. Therefore, we consider a new research scene for source-free unsupervised domain adaptation (SFUDA), which exclusively relies on a source model trained on source domain sample without requiring access to fully labeled source domain data. This paper introduces a SFUDA approach that utilizes knowledge distillation (KD), which involves two stages: (1) generalizing the source model by applying domain augmentation techniques and LS methods that enhance the model’s potential to enhance its generalization capability; (2) adapting the target model using a KD framework to achieve knowledge migration; and in addition, mutual information structure regularization is added to consider the internal data structure, thus enhancing the model’s adaptability. To evaluate the efficacy of our approach, we perform experiments on two datasets—the Case Western Reserve University dataset and the Paderborn University dataset, comprising 24 transfer tasks. Our experiments demonstrate the effectiveness of the domain augmentation technique, mutual information regularization, and the proposed method.

EDRL: Entropy-guided disentangled representation learning for unsupervised domain adaptation in semantic segmentation

Background and ObjectiveDeep learning-based approaches are excellent at learning from large amounts of data, but can be poor at generalizing the learned knowledge to testing datasets with domain shift, i.e., when there exists distribution discrepancy between the training dataset (source domain) and the testing dataset (target domain). In this paper, we investigate unsupervised domain adaptation (UDA) techniques to train a cross-domain segmentation method which is robust to domain shift, eliminating the requirement of any annotations on the target domain. MethodsTo this end, we propose an Entropy-guided Disentangled Representation Learning, referred as EDRL, for UDA in semantic segmentation. Concretely, we synergistically integrate image alignment via disentangled representation learning with feature alignment via entropy-based adversarial learning into one network, which is trained end-to-end. We additionally introduce a dynamic feature selection mechanism via soft gating, which helps to further enhance the task-specific feature alignment. We validate the proposed method on two publicly available datasets: the CT-MR dataset and the multi-sequence cardiac MR (MS-CMR) dataset. ResultsOn both datasets, our method achieved better results than the state-of-the-art (SOTA) methods. Specifically, on the CT-MR dataset, our method achieved an average DSC of 84.8% when taking CT as the source domain and MR as the target domain, and an average DSC of 84.0% when taking MR as the source domain and CT as the target domain. ConclusionsResults from comprehensive experiments demonstrate the efficacy of the proposed EDRL model for cross-domain medical image segmentation.

Unsupervised domain adaptation with and without access to source data for estimating occupancy and recognizing activities in smart buildings

Energy-efficient buildings have gained increasing interest in the last decades as they provide optimal energy management. With the emergence of smart homes, many smart tools have been developed to optimize energy efficiency such as activities recognition (AR) and occupancy estimation (OE). The creation of these smart building tools may be disrupted by the scarcity of labeled data. Indeed, data labeling is a tedious and time-consuming task and, it can be very expensive for building objectives. However, labeled data scarcity can be solved by sharing knowledge from different domains using unsupervised domain adaptation techniques. Also, privacy issues can emerge and prevent the use of certain types of data that share residents’ attitudes. However, data privacy can be preserved by considering approaches that do not have direct access to original data. In this research, we provide a comparative analysis between unsupervised domain adaptation (UDA) methods, applied to the tasks of AR and OE, with and without direct access to source domain data. We have 6 adapted approaches for UDA with access to labeled source data: domain separation networks (DSN), cluster alignment with a teacher (CAT), CAT+ gradient reversal (RevGrad), CAT + robust RevGrad (rRevGrad), Auxiliary Target Domain-Oriented Classifier (ATDOC) with nearest centroid classifier (NC), and ATDOC with neighborhood aggregation (NA). Also, we have 6 adapted methods for UDA without labeled source data: confidence score weighting adaptation using joint model data structure (CoWA-JMDS), CoWA-JMDS without weights mixup, divide and contrast (DaC), attracting and dispersing (AaD), source hypothesis transfer with information maximization (SHOT-IM), and source hypothesis transfer with self-supervised pseudo-labeling (SHOT-Pseudo-labeling). All the considered methods have been tested on AR and OE datasets collected using ambient sensors. The comparative analysis made in this work has shown us impressive findings and has given great ideas about the type of approaches (with access or without access to the source data) that we should consider for real-world smart building applications.

SSDA3D: Semi-supervised Domain Adaptation for 3D Object Detection from Point Cloud

LiDAR-based 3D object detection is an indispensable task in advanced autonomous driving systems. Though impressive detection results have been achieved by superior 3D detectors, they suffer from significant performance degeneration when facing unseen domains, such as different LiDAR configurations, different cities, and weather conditions. The mainstream approaches tend to solve these challenges by leveraging unsupervised domain adaptation (UDA) techniques. However, these UDA solutions just yield unsatisfactory 3D detection results when there is a severe domain shift, e.g., from Waymo (64-beam) to nuScenes (32-beam). To address this, we present a novel Semi-Supervised Domain Adaptation method for 3D object detection (SSDA3D), where only a few labeled target data is available, yet can significantly improve the adaptation performance. In particular, our SSDA3D includes an Inter-domain Adaptation stage and an Intra-domain Generalization stage. In the first stage, an Inter-domain Point-CutMix module is presented to efficiently align the point cloud distribution across domains. The Point-CutMix generates mixed samples of an intermediate domain, thus encouraging to learn domain-invariant knowledge. Then, in the second stage, we further enhance the model for better generalization on the unlabeled target set. This is achieved by exploring Intra-domain Point-MixUp in semi-supervised learning, which essentially regularizes the pseudo label distribution. Experiments from Waymo to nuScenes show that, with only 10% labeled target data, our SSDA3D can surpass the fully-supervised oracle model with 100% target label. Our code is available at https://github.com/yinjunbo/SSDA3D.

Distributionally robust unsupervised domain adaptation

Obtaining ground-truth label information from real-world data along with uncertainty quantification can be difficult or even infeasible. In the absence of labeled data for a certain task, unsupervised domain adaptation (UDA) techniques have shown great accomplishment by learning transferable knowledge from labeled source domain data and applying it to unlabeled target domain data, yet scarce studies consider addressing uncertainties under domain shifts to improve the model robustness. Distributionally robust learning (DRL) is emerging as a high-potential technique for building reliable learning systems that are robust to distribution shifts. In this study, we propose a distributionally robust unsupervised domain adaptation (DRUDA) method to enhance the machine learning model generalization ability under input space perturbations. The DRL-based UDA learning scheme is formulated as a min–max optimization problem by optimizing worst-case perturbations of the training source data. Our Wasserstein distributionally robust framework can reduce the shifts in the joint distributions across domains. The proposed DRUDA has been tested on digit datasets and the Office-31 dataset, compared with other state-of-the-art domain adaptation techniques. Our experimental results show that the proposed DRUDA leads to improvements in domain adaptation accuracy performance on target domains.

Domain shifts in dermoscopic skin cancer datasets: Evaluation of essential limitations for clinical translation

The limited ability of Convolutional Neural Networks to generalize to images from previously unseen domains is a major limitation, in particular, for safety-critical clinical tasks such as dermoscopic skin cancer classification. In order to translate CNN-based applications into the clinic, it is essential that they are able to adapt to domain shifts. Such new conditions can arise through the use of different image acquisition systems or varying lighting conditions. In dermoscopy, shifts can also occur as a change in patient age or occurrence of rare lesion localizations (e.g. palms). These are not prominently represented in most training datasets and can therefore lead to a decrease in performance. In order to verify the generalizability of classification models in real world clinical settings it is crucial to have access to data which mimics such domain shifts. To our knowledge no dermoscopic image dataset exists where such domain shifts are properly described and quantified. We therefore grouped publicly available images from ISIC archive based on their metadata (e.g. acquisition location, lesion localization, patient age) to generate meaningful domains. To verify that these domains are in fact distinct, we used multiple quantification measures to estimate the presence and intensity of domain shifts. Additionally, we analyzed the performance on these domains with and without an unsupervised domain adaptation technique. We observed that in most of our grouped domains, domain shifts in fact exist. Based on our results, we believe these datasets to be helpful for testing the generalization capabilities of dermoscopic skin cancer classifiers.

NuSegDA: Domain adaptation for nuclei segmentation.

The accurate segmentation of nuclei is crucial for cancer diagnosis and further clinical treatments. To successfully train a nuclei segmentation network in a fully-supervised manner for a particular type of organ or cancer, we need the dataset with ground-truth annotations. However, such well-annotated nuclei segmentation datasets are highly rare, and manually labeling an unannotated dataset is an expensive, time-consuming, and tedious process. Consequently, we require to discover a way for training the nuclei segmentation network with unlabeled dataset. In this paper, we propose a model named NuSegUDA for nuclei segmentation on the unlabeled dataset (target domain). It is achieved by applying Unsupervised Domain Adaptation (UDA) technique with the help of another labeled dataset (source domain) that may come from different type of organ, cancer, or source. We apply UDA technique at both of feature space and output space. We additionally utilize a reconstruction network and incorporate adversarial learning into it so that the source-domain images can be accurately translated to the target-domain for further training of the segmentation network. We validate our proposed NuSegUDA on two public nuclei segmentation datasets, and obtain significant improvement as compared with the baseline methods. Extensive experiments also verify the contribution of newly proposed image reconstruction adversarial loss, and target-translated source supervised loss to the performance boost of NuSegUDA. Finally, considering the scenario when we have a small number of annotations available from the target domain, we extend our work and propose NuSegSSDA, a Semi-Supervised Domain Adaptation (SSDA) based approach.

Adversarial Unsupervised Domain Adaptation for Hand Gesture Recognition Using Thermal Images

Hand gesture recognition has a wide range of applications, including in the automotive and industrial sectors, health assistive systems, authentication, and so on. Thermal images are more resistant to environmental changes than red–green–blue (RGB) images for hand gesture recognition. However, one disadvantage of using thermal images for the aforementioned task is the scarcity of labeled thermal datasets. To tackle this problem, we propose a method that combines unsupervised domain adaptation (UDA) techniques with deep-learning (DL) technology to remove the need for labeled data in the learning process. There are several types and methods for implementing UDA, with adversarial UDA being one of the most common. In this article, the first time in this field, we propose a novel adversarial UDA model that uses channel attention and bottleneck layers to learn domain-invariant features across RGB and thermal domains. Thus, the proposed model leverages the information from the labeled RGB data to solve the hand gesture recognition task using thermal images. We evaluate the proposed model on two hand gesture datasets, namely, Sign Digit Classification and Alphabet Gesture Classification, and compare it to other benchmark models in terms of accuracy, model size, and model parameters. Our model outperforms the other state-of-the-art methods on the Sign Digit Classification and Alphabet Gesture Classification datasets and achieves 91.32% and 80.91% target test accuracy, respectively.

Decompose to Adapt: Cross-Domain Object Detection Via Feature Disentanglement

Recent advances in unsupervised domain adaptation (UDA) techniques have witnessed great success in cross-domain computer vision tasks, enhancing the generalization ability of data-driven deep learning architectures by bridging the domain distribution gaps. For the UDA-based cross-domain object detection methods, the majority of them alleviate the domain bias by inducing the domain-invariant feature generation via adversarial learning strategy. However, their domain discriminators have limited classification ability due to the unstable adversarial training process. Therefore, the extracted features induced by them cannot be perfectly domain-invariant and still contain domain-private factors, bringing obstacles to further alleviate the cross-domain discrepancy. To tackle this issue, we design a Domain Disentanglement Faster-RCNN (DDF) to eliminate the source-specific information in the features for detection task learning. Our DDF method facilitates the feature disentanglement at the global and local stages, with a Global Triplet Disentanglement (GTD) module and an Instance Similarity Disentanglement (ISD) module, respectively. By outperforming state-of-the-art methods on four benchmark UDA object detection tasks, our DDF method is demonstrated to be effective with wide applicability.

Unsupervised Grammatical Correction With Optimized Layer Normalization and Dynamic Embedding Enhancement

Grammatical error correction aims to detect and correct grammatical errors with all types of mistaken, disordered, missing, and redundant characters. However, most existing methods focus more on detecting errors than correcting them. This paper proposes a domain-adaptive model with Interoperable Layer Normalization (ILN) and dynamic word embedding enhancement to optimize the error correction capability. To further improve the Chinese correction capability, we introduce multiple rounds of error correction to refine the sequence tagging model’s ability to fix mistakes. In addition, we propose a data augmentation method based on the complex tag to represent textual error correction traces more completely. We also explore a migration training method based on multiple training datasets. Further, we offer a unique unsupervised domain adaptation technique based on ILN, an innovative channel fusion approach that can significantly improve models’ domain adaptability. Finally, experimental results show that our proposed method substantially outperforms all robust baseline methods and achieves the best results in position-level and correction-level errors on the CGED-2020 dataset.

Unsupervised Adversarial Domain Adaptation for Agricultural Land Extraction of Remote Sensing Images

Agricultural land extraction is an essential technical means to promote sustainable agricultural development and modernization research. Existing supervised algorithms rely on many finely annotated remote-sensing images, which is both time-consuming and expensive. One way to reduce the annotation cost approach is to migrate models trained on existing annotated data (source domain) to unannotated data (target domain). However, model generalization capability is often unsatisfactory due to the limit of the domain gap. In this work, we use an unsupervised adversarial domain adaptation method to train a neural network to close the gap between the source and target domains for unsupervised agricultural land extraction. The overall approach consists of two phases: inter-domain and intra-domain adaptation. In the inter-domain adaptation, we use a generative adversarial network (GAN) to reduce the inter-domain gap between the source domain (labeled dataset) and the target domain (unlabeled dataset). The transformer with robust long-range dependency modeling acts as the backbone of the generator. In addition, the multi-scale feature fusion (MSFF) module is designed in the generator to accommodate remote sensing datasets with different spatial resolutions. Further, we use an entropy-based approach to divide the target domain. The target domain is divided into two subdomains, easy split images and hard split images. By training against each other between the two subdomains, we reduce the intra-domain gap. Experiments results on the “DeepGlobe → LoveDA”, “GID → LoveDA” and “DeepGlobe → GID” unsupervised agricultural land extraction tasks demonstrate the effectiveness of our method and its superiority to other unsupervised domain adaptation techniques.

A deep learning model for detection of Alzheimer's disease based on retinal photographs: a retrospective, multicentre case-control study

There is no simple model to screen for Alzheimer's disease, partly because the diagnosis of Alzheimer's disease itself is complex-typically involving expensive and sometimes invasive tests not commonly available outside highly specialised clinical settings. We aimed to develop a deep learning algorithm that could use retinal photographs alone, which is the most common method of non-invasive imaging the retina to detect Alzheimer's disease-dementia. In this retrospective, multicentre case-control study, we trained, validated, and tested a deep learning algorithm to detect Alzheimer's disease-dementia from retinal photographs using retrospectively collected data from 11 studies that recruited patients with Alzheimer's disease-dementia and people without disease from different countries. Our main aim was to develop a bilateral model to detect Alzheimer's disease-dementia from retinal photographs alone. We designed and internally validated the bilateral deep learning model using retinal photographs from six studies. We used the EfficientNet-b2 network as the backbone of the model to extract features from the images. Integrated features from four retinal photographs (optic nerve head-centred and macula-centred fields from both eyes) for each individual were used to develop supervised deep learning models and equip the network with unsupervised domain adaptation technique, to address dataset discrepancy between the different studies. We tested the trained model using five other studies, three of which used PET as a biomarker of significant amyloid β burden (testing the deep learning model between amyloid β positive vs amyloid β negative). 12 949 retinal photographs from 648 patients with Alzheimer's disease and 3240 people without the disease were used to train, validate, and test the deep learning model. In the internal validation dataset, the deep learning model had 83·6% (SD 2·5) accuracy, 93·2% (SD 2·2) sensitivity, 82·0% (SD 3·1) specificity, and an area under the receiver operating characteristic curve (AUROC) of 0·93 (0·01) for detecting Alzheimer's disease-dementia. In the testing datasets, the bilateral deep learning model had accuracies ranging from 79·6% (SD 15·5) to 92·1% (11·4) and AUROCs ranging from 0·73 (SD 0·24) to 0·91 (0·10). In the datasets with data on PET, the model was able to differentiate between participants who were amyloid β positive and those who were amyloid β negative: accuracies ranged from 80·6 (SD 13·4%) to 89·3 (13·7%) and AUROC ranged from 0·68 (SD 0·24) to 0·86 (0·16). In subgroup analyses, the discriminative performance of the model was improved in patients with eye disease (accuracy 89·6% [SD 12·5%]) versus those without eye disease (71·7% [11·6%]) and patients with diabetes (81·9% [SD 20·3%]) versus those without the disease (72·4% [11·7%]). A retinal photograph-based deep learning algorithm can detect Alzheimer's disease with good accuracy, showing its potential for screening Alzheimer's disease in a community setting. BrightFocus Foundation.