Local Dataset Research Articles

Collaborative learning from distributed data with differentially private synthetic data

BackgroundConsider a setting where multiple parties holding sensitive data aim to collaboratively learn population level statistics, but pooling the sensitive data sets is not possible due to privacy concerns and parties are unable to engage in centrally coordinated joint computation. We study the feasibility of combining privacy preserving synthetic data sets in place of the original data for collaborative learning on real-world health data from the UK Biobank.MethodsWe perform an empirical evaluation based on an existing prospective cohort study from the literature. Multiple parties were simulated by splitting the UK Biobank cohort along assessment centers, for which we generate synthetic data using differentially private generative modelling techniques. We then apply the original study’s Poisson regression analysis on the combined synthetic data sets and evaluate the effects of 1) the size of local data set, 2) the number of participating parties, and 3) local shifts in distributions, on the obtained likelihood scores.ResultsWe discover that parties engaging in the collaborative learning via shared synthetic data obtain more accurate estimates of the regression parameters compared to using only their local data. This finding extends to the difficult case of small heterogeneous data sets. Furthermore, the more parties participate, the larger and more consistent the improvements become up to a certain limit. Finally, we find that data sharing can especially help parties whose data contain underrepresented groups to perform better-adjusted analysis for said groups.ConclusionsBased on our results we conclude that sharing of synthetic data is a viable method for enabling learning from sensitive data without violating privacy constraints even if individual data sets are small or do not represent the overall population well. Lack of access to distributed sensitive data is often a bottleneck in biomedical research, which our study shows can be alleviated with privacy-preserving collaborative learning methods.

Parallel processing model for low-dose computed tomography image denoising

Low-dose computed tomography (LDCT) has gained increasing attention owing to its crucial role in reducing radiation exposure in patients. However, LDCT-reconstructed images often suffer from significant noise and artifacts, negatively impacting the radiologists’ ability to accurately diagnose. To address this issue, many studies have focused on denoising LDCT images using deep learning (DL) methods. However, these DL-based denoising methods have been hindered by the highly variable feature distribution of LDCT data from different imaging sources, which adversely affects the performance of current denoising models. In this study, we propose a parallel processing model, the multi-encoder deep feature transformation network (MDFTN), which is designed to enhance the performance of LDCT imaging for multisource data. Unlike traditional network structures, which rely on continual learning to process multitask data, the approach can simultaneously handle LDCT images within a unified framework from various imaging sources. The proposed MDFTN consists of multiple encoders and decoders along with a deep feature transformation module (DFTM). During forward propagation in network training, each encoder extracts diverse features from its respective data source in parallel and the DFTM compresses these features into a shared feature space. Subsequently, each decoder performs an inverse operation for multisource loss estimation. Through collaborative training, the proposed MDFTN leverages the complementary advantages of multisource data distribution to enhance its adaptability and generalization. Numerous experiments were conducted on two public datasets and one local dataset, which demonstrated that the proposed network model can simultaneously process multisource data while effectively suppressing noise and preserving fine structures. The source code is available at https://github.com/123456789ey/MDFTN.

Local protest event analysis: providing a more comprehensive picture?

In exploring protest dynamics, Protest Event Analysis (PEA) has proven an indispensable analytic tool. Despite various improvements, PEA faces continuous challenges, notably the reproduction of media-specific selection biases. This research note aims to contribute to the literature seeking to mitigate these issues by exploring the potential of PEA based on local newspaper data, which tend to be less selective in their protest reporting. The article introduces an original local PEA dataset in Germany (2000–2020) and systematically compares it to existing national-level PEA data for Germany. The analysis shows that while both datasets pick up similar trends in terms of mobilisation waves, national level data focus on shorter periods of intense protest activity. Furthermore, issues of protest differ: material and social issues (including labour and environmental issues) are more prominent in local data, while cultural issues (including immigration and gender) are more prominent in national-level data.

Federated Learning in the Detection of Fake News Using Deep Learning as a Basic Method.

This article explores the possibilities for federated learning with a deep learning method as a basic approach to train detection models for fake news recognition. Federated learning is the key issue in this research because this kind of learning makes machine learning more secure by training models on decentralized data at decentralized places, for example, at different IoT edges. The data are not transformed between decentralized places, which means that personally identifiable data are not shared. This could increase the security of data from sensors in intelligent houses and medical devices or data from various resources in online spaces. Each station edge could train a model separately on data obtained from its sensors and on data extracted from different sources. Consequently, the models trained on local data on local clients are aggregated at the central ending point. We have designed three different architectures for deep learning as a basis for use within federated learning. The detection models were based on embeddings, CNNs (convolutional neural networks), and LSTM (long short-term memory). The best results were achieved using more LSTM layers (F1 = 0.92). On the other hand, all three architectures achieved similar results. We also analyzed results obtained using federated learning and without it. As a result of the analysis, it was found that the use of federated learning, in which data were decomposed and divided into smaller local datasets, does not significantly reduce the accuracy of the models.

A Semi-supervised Four-Chamber Echocardiographic Video Segmentation Algorithm Based on Multilevel Edge Perception and Calibration Fusion

ObjectiveEchocardiographic videos are commonly used for automatic semantic segmentation of endocardium, which is crucial in evaluating cardiac function and assisting doctors to make accurate diagnoses of heart disease. However, this task faces two distinct challenges: one is the edge blurring, which is caused by the presence of speckle noise or excessive de-noising operation, and the other is the lack of an effective feature fusion approach for multilevel features for obtaining accurate endocardium. MethodsIn this study, a deep learning model, based on multilevel edge perception and calibration fusion is proposed to improve the segmentation performance. First, a multilevel edge perception module is proposed to comprehensively extract edge features through both a detail branch and a semantic branch to alleviate the adverse impact of noise. Second, a calibration fusion module is proposed that calibrates and integrates various features, including semantic and detailed information, to maximize segmentation performance. Furthermore, the features obtained from the calibration fusion module are stored by using a memory architecture to achieve semi-supervised segmentation through both labeled and unlabeled data. ResultsOur method is evaluated on two public echocardiography video data sets, achieving average Dice coefficients of 93.05% and 93.93%, respectively. Additionally, we validated our method on a local hospital clinical data set, achieving a Pearson correlation of 0.765 for predicting left ventricular ejection fraction. ConclusionThe proposed model effectively solves the challenges encountered in echocardiography by using semi-supervised networks, thereby improving the segmentation accuracy of the ventricles. This indicates that the proposed model can assist cardiologists in obtaining accurate and effective research and diagnostic results.

Explainable Federated Medical Image Analysis Through Causal Learning and Blockchain.

Federated learning (FL) enables collaborative training of machine learning models across distributed medical data sources without compromising privacy. However, applying FL to medical image analysis presents challenges like high communication overhead and data heterogeneity. This paper proposes novel FL techniques using explainable artificial intelligence (XAI) for efficient, accurate, and trustworthy analysis. A heterogeneity-aware causal learning approach selectively sparsifies model weights based on their causal contributions, significantly reducing communication requirements while retaining performance and improving interpretability. Furthermore, blockchain provides decentralized quality assessment of client datasets. The assessment scores adjust aggregation weights so higher-quality data has more influence during training, improving model generalization. Comprehensive experiments show our XAI-integrated FL framework enhances efficiency, accuracy and interpretability. The causal learning method decreases communication overhead while maintaining segmentation accuracy. The blockchain-based data valuation mitigates issues from low-quality local datasets. Our framework provides essential model explanations and trust mechanisms, making FL viable for clinical adoption in medical image analysis.

Expansive Receptive Field and Local Feature Extraction Network: Advancing Multiscale Feature Fusion for Breast Fibroadenoma Segmentation in Sonography.

Fibroadenoma is a common benign breast disease that affects women of all ages. Early diagnosis can greatly improve the treatment outcomes and reduce the associated pain. Computer-aided diagnosis (CAD) has great potential to improve diagnosis accuracy and efficiency. However, its application in sonography is limited. A network that utilizes expansive receptive fields and local information learning was proposed for the accurate segmentation of breast fibroadenomas in sonography. The architecture comprises the Hierarchical Attentive Fusion module, which conducts local information learning through channel-wise and pixel-wise perspectives, and the Residual Large-Kernel module, which utilizes multiscale large kernel convolution for global information learning. Additionally, multiscale feature fusion in both modules was included to enhance the stability of our network. Finally, an energy function and a data augmentation method were incorporated to fine-tune low-level features of medical images and improve data enhancement. The performance of our model is evaluated using both our local clinical dataset and a public dataset. Mean pixel accuracy (MPA) of 93.93% and 86.06% and mean intersection over union (MIOU) of 88.16% and 73.19% were achieved on the clinical and public datasets, respectively. They are significantly improved over state-of-the-art methods such as SegFormer (89.75% and 78.45% in MPA and 83.26% and 71.85% in MIOU, respectively). The proposed feature extraction strategy, combining local pixel-wise learning with an expansive receptive field for global information perception, demonstrates excellent feature learning capabilities. Due to this powerful and unique local-global feature extraction capability, our deep network achieves superior segmentation of breast fibroadenoma in sonography, which may be valuable in early diagnosis.

Privacy-preserving intelligent fault diagnostics for wind turbine clusters using federated stacked capsule autoencoder

The emergence of Internet of Things (IoT) technologies in the field of health monitoring has introduced the paradigm of Industrial Internet of Things (IIoT) to the industry. IIoT systems provide enterprises with a substantial volume of monitoring data for industrial equipment health monitoring, facilitating the development of artificial intelligence fault diagnosis models. However, a singular industrial entity often encounters limitations in collecting sufficient training data in practical scenarios. Moreover, the sharing of confidential information among entities is strictly prohibited due to concerns regarding intellectual property and data security. This study proposes a fault diagnosis system that addresses this issue by incorporating a capsule-based fault feature expression into the federated learning (FL) framework. The system comprises clients distributed across multiple factories and a central server hosted in the cloud. The client models are trained on local private datasets, and then knowledge fusion is achieved by uploading intrinsic templates and pose matrices to the central server. The proposed method offers the advantage of reducing transmission burden and enhancing data security in comparison to existing FL approaches. Besides, a capsule knowledge alignment algorithm is proposed to update the capsule-based fault feature expression ona central server. To simulate real fault diagnosis application scenarios, two similar fault simulation platforms are built to acquire isolated fault diagnosis datasets. The effectiveness of the proposed method is verified using these datasets.

A global map of gullied hillslopes on Mars

The distribution of gullies on Mars has been used to support different modes of formation involving CO2, H2O or entirely dry processes. We mapped the extent of gullied hillslope using a global mosaic of ConTeXt camera (CTX) at 6 m/pixel and High Resolution Imaging Science Experiment (HiRISE) browse products at ∼3 m/pixel. Our new global catalog of gullied hillslopes allows for quantitative, area-based analysis of gullies both at global and regional scale, which was not possible with point-based or more localized previous datasets. We analyzed aspect and gradient information derived from MOLA topography for every mapped gullied hillslope. We have confirmed the trends identified in previous works, which are that the mean gradient of gullied hillslopes is lower than 30°, the angle of repose of dry material on Mars, and that the preferred aspect of gullied hillslopes is latitude-dependent, with a shift between pole-facing to equator-facing happening around 40° in both hemispheres. We established a hierarchy of factors that can explain the global distribution of gullied hillslopes: Latitude is the strictest control because gullies are constrained between 26°-83°S and 28°-76°N. The availability of steep slopes is a strong control over gully distribution, and we observed that gullied hillslopes aspect reflects the orientation of large regional reliefs, which implies that hillslope aspect is less of a factor than the hillslope gradient in explaining gully distribution. We analyzed the correlation of gully distribution with a subsurface water ice consistency map, and we found that gullies are preferentially distributed on areas with some ground ice rather than no ice at all. We identified for the first time a bias toward an east-facing trend for southern hemisphere gullies poleward of ∼40° of latitude. We also linked our gullied hillslopes dataset with a global crater morphometry dataset: 63% of the gullied area on Mars is found in craters, and we find that the area covered by gullies in a given crater is positively correlated with the depth and diameter ratio of this crater.Our observations point toward a complex interplay between local slope, insolation, and thermal properties of the substrate to explain the spatial distribution of the current population of gullies. Finally, we have assessed the very high resolution (sub-meter) imagery coverage of gullies to be 36% and that the repeated coverage reaches 21% of gullies, hence this would need expanding to better understand the spatial trends in active gullies.

An Enhanced Mask R-CNN Approach for Pulmonary Embolism Detection and Segmentation.

Pulmonary embolism (PE) refers to the occlusion of pulmonary arteries by blood clots, posing a mortality risk of approximately 30%. The detection of pulmonary embolism within segmental arteries presents greater challenges compared with larger arteries and is frequently overlooked. In this study, we developed a computational method to automatically identify pulmonary embolism within segmental arteries using computed tomography (CT) images. The system architecture incorporates an enhanced Mask R-CNN deep neural network trained on PE-containing images. This network accurately localizes pulmonary embolisms in CT images and effectively delineates their boundaries. This study involved creating a local data set and evaluating the model predictions against pulmonary embolisms manually identified by expert radiologists. The sensitivity, specificity, accuracy, Dice coefficient, and Jaccard index values were obtained as 96.2%, 93.4%, 96.%, 0.95, and 0.89, respectively. The enhanced Mask R-CNN model outperformed the traditional Mask R-CNN and U-Net models. This study underscores the influence of Mask R-CNN's loss function on model performance, providing a basis for the potential improvement of Mask R-CNN models for object detection and segmentation tasks in CT images.

Anomaly Detection and Biomarkers Localization in Retinal Images.

Background: To design a novel anomaly detection and localization approach using artificial intelligence methods using optical coherence tomography (OCT) scans for retinal diseases. Methods: High-resolution OCT scans from the publicly available Kaggle dataset and a local dataset were used by four state-of-the-art self-supervised frameworks. The backbone model of all the frameworks was a pre-trained convolutional neural network (CNN), which enabled the extraction of meaningful features from OCT images. Anomalous images included choroidal neovascularization (CNV), diabetic macular edema (DME), and the presence of drusen. Anomaly detectors were evaluated by commonly accepted performance metrics, including area under the receiver operating characteristic curve, F1 score, and accuracy. Results: A total of 25,315 high-resolution retinal OCT slabs were used for training. Test and validation sets consisted of 968 and 4000 slabs, respectively. The best performing across all anomaly detectors had an area under the receiver operating characteristic of 0.99. All frameworks were shown to achieve high performance and generalize well for the different retinal diseases. Heat maps were generated to visualize the quality of the frameworks' ability to localize anomalous areas of the image. Conclusions: This study shows that with the use of pre-trained feature extractors, the frameworks tested can generalize to the domain of retinal OCT scans and achieve high image-level ROC-AUC scores. The localization results of these frameworks are promising and successfully capture areas that indicate the presence of retinal pathology. Moreover, such frameworks have the potential to uncover new biomarkers that are difficult for the human eye to detect. Frameworks for anomaly detection and localization can potentially be integrated into clinical decision support and automatic screening systems that will aid ophthalmologists in patient diagnosis, follow-up, and treatment design. This work establishes a solid basis for further development of automated anomaly detection frameworks for clinical use.

SuiETAS: Developing and Testing ETAS-Based Earthquake Forecasting Models for Switzerland

ABSTRACT We present the development and testing of multiple epidemic-type aftershock sequence (ETAS)-based earthquake forecasting models for Switzerland, aiming to identify suitable candidate models for operational earthquake forecasting (OEF) at the Swiss Seismological Service. We consider seven model variants: four variants use parameters obtained through fitting the ETAS model to the Swiss earthquake catalog, and three use generic parameters that were fit to Californian seismicity or global seismicity from regions tectonically similar to Switzerland. The model variants use different pieces of information from the current state-of-the-art time-independent earthquake rate forecast underlying the Swiss seismic hazard model (SUIhaz2015), and one is calibrated on a larger local data set that includes smaller earthquakes by allowing a time-dependent estimate of the completeness magnitude. We test all variants using pseudoprospective short-term (7-day) forecasting experiments and retrospective long-term (30-year) consistency tests. Our results suggest that all ETAS-based models outperform the time-independent SUIhaz2015 forecast in the short term, but two of the model variants overestimate event numbers in the long term. ETAS parameters are found not to be universally transferrable across tectonic regimes, and region-specific calibration is found to add value over generic parameters. Finally, we rank all model variants based on six criteria, including the results of the pseudoprospective and retrospective tests, as well as other criteria such as model run time or consistency with the existing long-term model, using a multicriteria decision analysis approach. Based on this ranking, we propose the ETAS model calibrated on the Swiss catalog, and with the spatial background seismicity distribution of SUIhaz2015 as the ideal candidate for the first Swiss OEF system. All procedures and choices involved in the development and testing of the Swiss ETAS model follow recently established expert recommendations and can act as a reference in the establishment of time-variant earthquake forecasting models for other regions.

Toward drift-free high-throughput nanoscopy through adaptive intersection maximization.

Single-molecule localization microscopy (SMLM) often suffers from suboptimal resolution due to imperfect drift correction. Existing marker-free drift correction algorithms often struggle to reliably track high-frequency drift and lack the computational efficiency to manage large, high-throughput localization datasets. We present an adaptive intersection maximization-based method (AIM) that leverages the entire dataset's information content to minimize drift correction errors, particularly addressing high-frequency drift, thereby enhancing the resolution of existing SMLM systems. We demonstrate that AIM can robustly and efficiently achieve an angstrom-level tracking precision for high-throughput SMLM datasets under various imaging conditions, resulting in an optimal resolution in simulated and biological experimental datasets. We offer AIM as one simple, model-free software for instant resolution enhancement with standard CPU devices.

A Contribution-Aware Noise Feature representation model for image manipulation localization

Noise feature modules play a prevalent role in image manipulation localization. However, existing approach involves utilizing a specific noise feature module tailored to address distinct categories of tampering procedures, limiting its adaptability across diverse scenarios. This limitation becomes particularly pronounced in image manipulation localization where a singular noise module struggles to comprehensively represent all potential tampering types, given the inherent uncertainty surrounding the nature of tampering. A novel Contribution-Aware Noise Feature (CANF) representation model, specifically designed to tackle the intricacies of image tamper noise, is introduced in this paper. The proposed model integrates existing noise feature modules, autonomously enhancing their capabilities. This enhancement not only contributes significantly to image manipulation localization during the inference phase but also enables a quantitative evaluation of the distinct contributions of various noise modules to the inference results. To optimize the utilization of image information, the Red–Green–Blue (RGB) and CANF data are fused, leading to the development of a standardized image manipulation localization architecture. This architecture exhibits heightened efficiency in harnessing diverse types of information, thereby enhancing the overall performance of image manipulation localization. Experimental results demonstrate that our approach, alongside established methods IF-OSN, MSFF, PCL, and NCL, surpasses the state-of-the-art NCL by 8.5% and 5.4% concerning average F1 and AUC metrics across four standard image manipulation localization datasets: NIST16, CASIAv1, Columbia, and Coverage. In addition, we comprehensively explore the impact of various noise characterization modules on different types of tampering and the fact that Late Fusion is more conducive to image tampering localization.

PrivatEyes: Appearance-based Gaze Estimation Using Federated Secure Multi-Party Computation

Latest gaze estimation methods require large-scale training data but their collection and exchange pose significant privacy risks. We propose PrivatEyes - the first privacy-enhancing training approach for appearance-based gaze estimation based on federated learning (FL) and secure multi-party computation (MPC). PrivatEyes enables training gaze estimators on multiple local datasets across different users and server-based secure aggregation of the individual estimators' updates. PrivatEyes guarantees that individual gaze data remains private even if a majority of the aggregating servers is malicious. We also introduce a new data leakage attack DualView that shows that PrivatEyes limits the leakage of private training data more effectively than previous approaches. Evaluations on the MPIIGaze, MPIIFaceGaze, GazeCapture, and NVGaze datasets further show that the improved privacy does not lead to a lower gaze estimation accuracy or substantially higher computational costs - both of which are on par with its non-secure counterparts.

Advancing Federated Learning by Addressing Data and System Heterogeneity

In the emerging field of federated learning (FL), the challenge of heterogeneity, both in data and systems, presents significant obstacles to efficient and effective model training. This talk focuses on the latest advancements and solutions addressing these challenges. The first part of the talk delves into data heterogeneity, a core issue in FL, where data distributions across different clients vary widely and affect FL convergence. We will introduce the FedCor framework addressing this by modeling loss correlations between clients using Gaussian Process and reducing expected global loss. External covariate shift in FL is uncovered, demonstrating that normalization layers are crucial, and layer normalization proves effective. Additionally, class imbalance in FL degrades performance, but our proposed Federated Class-balanced Sampling (Fed-CBS) mechanism reduces this imbalance by employing homomorphic encryption for privacy preservation. The second part of the talk shifts focus to system heterogeneity, an equally critical challenge in FL. System heterogeneity involves the varying computational capabilities, network speeds, and other resource-related constraints of participating devices in FL. To address this, we introduce FedSEA, which is a semi-asynchronous FL framework that addresses accuracy drops by balancing aggregation frequency and predicting local update arrival. Additionally, we discuss FedRepre, a framework specifically designed to enhance FL in real-world environments by addressing challenges including unbalanced local dataset distributions, uneven computational capabilities, and fluctuating network speeds. By introducing a client selection mechanism and a specialized server architecture, FedRepre notably improves the efficiency, scalability, and performance of FL systems. Our talk aims to provide a comprehensive overview of the current research and advancements in tackling both data and system heterogeneity in federated learning. We hope to highlight the path forward for FL, underlining its potential in diverse real-world applications while maintaining data privacy and optimizing resource usage.

ADHD classification with cross-dataset feature selection for biomarker consistency detection

Objective. Attention deficit hyperactivity disorder (ADHD) is a prevalent neurodevelopmental disorder in children. While numerous intelligent methods are applied for its subjective diagnosis, they seldom consider the consistency problem of ADHD biomarkers. In practice, these data-driven approaches lead to varying learned features for ADHD classification across diverse ADHD datasets. This phenomenon significantly undermines the reliability of identified biomarkers and hampers the interpretability of these methods. Approach. In this study, we propose a cross-dataset feature selection (FS) module using a grouped SVM-based recursive feature elimination approach (G-SVM-RFE) to enhance biomarker consistency across multiple datasets. Additionally, we employ connectome gradient data for ADHD classification. In details, we introduce the G-SVM-RFE method to effectively concentrate gradient components within a few brain regions, thereby increasing the likelihood of identifying these regions as ADHD biomarkers. The cross-dataset FS module is integrated into an existing binary hypothesis testing (BHT) framework. This module utilizes external datasets to identify global regions that yield stable biomarkers. Meanwhile, given a dataset which waits for implementing the classification task as local dataset, we learn its own specific regions to further improve the performance of accuracy on this dataset. Main results. By employing this module, our experiments achieve an average accuracy of 96.7% on diverse datasets. Importantly, the discriminative gradient components primarily originate from the global regions, providing evidence for the significance of these regions. We further identify regions with the high appearance frequencies as biomarkers, where all the used global regions and one local region are recognized. Significance. These biomarkers align with existing research on impaired brain regions in children with ADHD. Thus, our method demonstrates its validity by providing enhanced biological explanations derived from ADHD mechanisms.

A Direct Approach for Local Quasi-Geoid Modeling Based on Spherical Radial Basis Functions Using a Noisy Satellite-Only Global Gravity Field Model

A Direct Approach for Local Quasi-Geoid Modeling Based on Spherical Radial Basis Functions Using a Noisy Satellite-Only Global Gravity Field Model

MedKnee: A New Deep Learning-Based Software for Automated Prediction of Radiographic Knee Osteoarthritis.

In computer-aided medical diagnosis, deep learning techniques have shown that it is possible to offer performance similar to that of experienced medical specialists in the diagnosis of knee osteoarthritis. In this study, a new deep learning (DL) software, called "MedKnee" is developed to assist physicians in the diagnosis process of knee osteoarthritis according to the Kellgren and Lawrence (KL) score. To accomplish this task, 5000 knee X-ray images obtained from the Osteoarthritis Initiative public dataset (OAI) were divided into train, valid, and test datasets in a ratio of 7:1:2 with a balanced distribution across each KL grade. The pre-trained Xception model is used for transfer learning and then deployed in a Graphical User Interface (GUI) developed with Tkinter and Python. The suggested software was validated on an external public database, Medical Expert, and compared with a rheumatologist's diagnosis on a local database, with the involvement of a radiologist for arbitration. The MedKnee achieved an accuracy of 95.36% when tested on Medical Expert-I and 94.94% on Medical Expert-II. In the local dataset, the developed tool and the rheumatologist agreed on 23 images out of 30 images (74%). The MedKnee's satisfactory performance makes it an effective assistant for doctors in the assessment of knee osteoarthritis.

Impact of Transfer Learning Using Local Data on Performance of a Deep Learning Model for Screening Mammography.

Purpose To investigate the issues of generalizability and replication of deep learning models by assessing performance of a screening mammography deep learning system developed at New York University (NYU) on a local Australian dataset. Materials and Methods In this retrospective study, all individuals with biopsy or surgical pathology-proven lesions and age-matched controls were identified from a South Australian public mammography screening program (January 2010 to December 2016). The primary outcome was deep learning system performance-measured with area under the receiver operating characteristic curve (AUC)-in classifying invasive breast cancer or ductal carcinoma in situ (n = 425) versus no malignancy (n = 490) or benign lesions (n = 44). The NYU system, including models without (NYU1) and with (NYU2) heatmaps, was tested in its original form, after training from scratch (without transfer learning), and after retraining with transfer learning. Results The local test set comprised 959 individuals (mean age, 62.5 years ± 8.5 [SD]; all female). The original AUCs for the NYU1 and NYU2 models were 0.83 (95% CI: 0.82, 0.84) and 0.89 (95% CI: 0.88, 0.89), respectively. When NYU1 and NYU2 were applied in their original form to the local test set, the AUCs were 0.76 (95% CI: 0.73, 0.79) and 0.84 (95% CI: 0.82, 0.87), respectively. After local training without transfer learning, the AUCs were 0.66 (95% CI: 0.62, 0.69) and 0.86 (95% CI: 0.84, 0.88). After retraining with transfer learning, the AUCs were 0.82 (95% CI: 0.80, 0.85) and 0.86 (95% CI: 0.84, 0.88). Conclusion A deep learning system developed using a U.S. dataset showed reduced performance when applied "out of the box" to an Australian dataset. Local retraining with transfer learning using available model weights improved model performance. Keywords: Screening Mammography, Convolutional Neural Network (CNN), Deep Learning Algorithms, Breast Cancer Supplemental material is available for this article. © RSNA, 2024 See also commentary by Cadrin-Chênevert in this issue.