Institute of High Performance Computing

White matter hyperintensities (WMH) are neuroimaging markers linked to an elevated risk of cognitive decline. WMH severity is typically assessed via visual rating scales and through volumetric segmentation. While visual rating scales are commonly used in clinical practice, they offer limited descriptive power. In contrast, supervised volumetric segmentation requires manually annotated masks, which are labor‐intensive and challenging to scale for large studies. Therefore, our goal was to develop an automated deep‐learning model that can provide accurate and holistic quantification of WMH severity with minimal supervision. We developed WMH‐DualTasker, a deep learning model that simultaneously performs voxel‐wise segmentation and visual rating score prediction. The model employs self‐supervised learning with transformation‐invariant consistency constraints, using WMH visual ratings (ARWMC scale, range 0–30) from clinical settings as the sole supervisory signal. Additionally, we assessed its clinical utility by applying it to identify individuals with mild cognitive impairment (MCI) and to predict dementia conversion. The volumetric quantification performance of WMH‐DualTasker was either superior to or on par with existing supervised methods, as demonstrated on the MICCAI‐WMH dataset (N = 60, Dice = 0.602) and the SINGER dataset (N = 64, Dice = 0.608). Furthermore, the model exhibited strong agreement with clinical visual rating scales on an external dataset (SINGER, MAE = 1.880, K = 0.77). Importantly, WMH severity metrics derived from WMH‐DualTasker improved predictive performance beyond conventional clinical features for MCI classification (AUC = 0.718, p < 0.001) and MCI conversion prediction (AUC = 0.652, p < 0.001) using the ADNI dataset. WMH‐DualTasker substantially reduces the reliance on labor‐intensive manual annotations, facilitating more efficient and scalable quantification of WMH severity in large‐scale population studies. This innovative approach has the potential to advance preventive and precision medicine by enhancing the assessment and management of vascular cognitive impairment associated with WMH.

Neural architecture search (NAS) is an automated technique to design optimal neural network architectures for a specific workload. Conventionally, evaluating candidate networks in NAS involves extensive training, which requires significant time and computational resources. To address this, training-free NAS has been proposed to expedite network evaluation with minimal search time. However, state-of-the-art training-free NAS algorithms struggle to precisely distinguish well-performing networks from poorly performing networks, resulting in inaccurate performance predictions and consequently suboptimal top-one network accuracy. Moreover, they are less effective in activation function exploration. To tackle the challenges, this article proposes RBFleX-NAS, a novel training-free NAS framework that accounts for both activation outputs and input features of the last layer with a radial basis function (RBF) kernel. We also present a detection algorithm to identify optimal hyperparameters using the obtained activation outputs and input feature maps. We verify the efficacy of RBFleX-NAS over a variety of NAS benchmarks. RBFleX-NAS significantly outperforms state-of-the-art training-free NAS methods in terms of top-one accuracy, achieving this with short search time in NAS-Bench-201 and NAS-Bench-SSS. In addition, it demonstrates a higher Kendall correlation compared to layer-based training-free NAS algorithms. Furthermore, we propose the neural network activation function benchmark (NAFBee), a new activation design space that extends the activation type to encompass various commonly used functions. In this extended design space, RBFleX-NAS demonstrates its superiority by accurately identifying the best-performing network during activation function search, providing a significant advantage over other NAS algorithms.

The rapid adoption of Artificial Intelligence (AI) in medical imaging raises fairness and privacy concerns across demographic groups, especially in diagnosis and treatment decisions. While federated learning (FL) offers decentralized privacy preservation, current frameworks often prioritize collaboration fairness over group fairness, risking healthcare disparities. Here we present FlexFair, an innovative FL framework designed to address both fairness and privacy challenges. FlexFair incorporates a flexible regularization term to facilitate the integration of multiple fairness criteria, including equal accuracy, demographic parity, and equal opportunity. Evaluated across four clinical applications (polyp segmentation, fundus vascular segmentation, cervical cancer segmentation, and skin disease diagnosis), FlexFair outperforms state-of-the-art methods in both fairness and accuracy. Moreover, we curate a multi-center dataset for cervical cancer segmentation that includes 678 patients from four hospitals. This diverse dataset allows for a more comprehensive analysis of model performance across different population groups, ensuring the findings are applicable to a broader range of patients.

Tuning of catalyst-support interactions potentially offers a powerful means to control activity. However, rational design of the catalyst support is challenged by a lack of clear property-activity relationships. Here, we uncover how the electronegativity of a support influences reaction pathways in electrochemical CO 2 reduction. This was achieved by creating a model system consisting of Cu nanoparticles hosted on a series of carbon supports, each with a different heteroatom dopant of varying electronegativity. Notably, we discovered that dopants with high electronegativity reduce the electron density on Cu and induce a selectivity shift toward multicarbon (C 2+ ) products. With this design principle, we built a composite Cu and F-doped carbon catalyst that achieves a C 2+ Faradaic efficiency of 82.5% at 400 mA cm ⁻² , with stable performance for 44 hours. Using simulated flue gas, the catalyst attains a C 2+ FE of 27.3%, which is a factor of 5.3 times higher than a reference Cu catalyst.

Past research on option generation, the mental process of creating possible courses of action for goal‐directed behaviors, focused extensively on the outcomes of the process, specifically, the quantity and quality of options generated. Accordingly, various effects were introduced to describe and categorize observed trends in option properties, yet these studies utilize differing task designs. This paper focuses on the “quantity‐breeds‐quality”, “less‐is‐more”, and the concomitant “Take The First” (TTF) heuristics. We conducted a secondary analysis of data from a culture‐free, education‐independent, and quantitative option generation task and compared the results to those predicted by the heuristics to discuss how study characteristics are well‐aligned with the heuristics they investigate. To bolster ecological validity and reflect a more diverse range of cognitive experiences beyond the neurotypical population, 44 healthy individuals and 54 patients with Major Depressive Disorder were asked to generate as many different paths as they could between two fixed points on a touchscreen computer in 1.5 min, and the generated options were quantified based on three metrics of interest: fluency, uniqueness, and diversity. For both groups, the mean uniqueness, maximum uniqueness, and diversity of an individual's paths were negatively correlated with an increase in fluency, in line with the less‐is‐more effect yet conflicting with the results predicted by the quantity‐breeds‐quality effect. In addition, normalized path uniqueness decreased with the path index, contrary to the results predicted by the TTF heuristic. The results were analyzed with reference to the three heuristics, to discuss possible task characteristics that cause a particular heuristic to apply, and demonstrate the fundamental differences between real‐life decision‐making scenarios and knowledge‐independent tasks.

Purpose To use artificial intelligence (AI) for quantifying schisis volume (ASV) in X‐linked retinoschisis (XLRS) for use as a structural endpoint in gene therapy clinical trials. Methods We used data from Singapore, the United Kingdom, the Netherlands, and the United States. The AI model was developed on 250 optical coherence tomography (OCT) slices, with human annotation of schisis cavities (Dataset 1). ASV was quantified on Dataset 2 – 16 OCT scans from 8 eyes with XLRS at two time points, and Dataset 4 – 62 OCT scans from 31 eyes at two time points before and after carbonic anhydrase inhibitor (CAI) treatment. A clinical trial was simulated comparing CAI treatment against control. Changes in ASV, central subfield thickness (CST) and central foveal thickness (CFT) were compared. Effect size (Cohen's d ) of the three structural endpoints was determined and used in sample size calculations for a future XLRS gene therapy clinical trial, at a 0.05 significance level and 80% power. Results In the simulated clinical trial, all structural metrics showed greater reductions with intervention than with control, but only change in ASV reached statistical significance ( p = 0.004). Cohen's d for ASV, CST and CFT were 0.972, 0.685 and 0.521, respectively. For the future gene therapy clinical trial, sample sizes required in each arm for ASV, CST and CFT were 18, 35 and 59 participants, respectively. Conclusions ASV measurements can track changes in schisis volume in response to treatment. As an endpoint, ASV has a greater statistical effect size than CST/CFT, which reduces sample size requirements for future XLRS gene therapy clinical trials.

Dataset distillation techniques have revolutionized the way of utilizing large datasets by compressing them into smaller, yet highly effective subsets that preserve the original datasets’ accuracy. However, while these methods have proven effective in reducing data size and training times, the robustness of these distilled datasets against adversarial attacks remains underexplored. This vulnerability poses significant risks, particularly in security-sensitive applications. To address this critical gap, we introduce DD-RobustBench, a novel and comprehensive benchmark specifically designed to evaluate the adversarial robustness of distilled datasets. Our benchmark is the most extensive of its kind and integrates a variety of dataset distillation techniques, including recent advancements such as TESLA, DREAM, SRe2L, and D4M, which have shown promise in enhancing model performance. DD-RobustBench also rigorously tests these datasets against a diverse array of adversarial attack methods to ensure broad applicability. Our evaluations cover a wide spectrum of datasets, including but not limited to, the widely used ImageNet-1K. This allows us to assess the robustness of distilled datasets in scenarios mirroring real-world applications. Furthermore, our detailed quantitative analysis investigates how different components involved in the distillation process, such as data augmentation, downsampling, and clustering, affect dataset robustness. Our findings provide critical insights into which techniques enhance or weaken the resilience of distilled datasets against adversarial threats, offering valuable guidelines for developing more robust distillation methods in the future. Through DD-RobustBench, we aim not only to benchmark but also to push the boundaries of dataset distillation research by highlighting areas for improvement and suggesting pathways for future innovations in creating datasets that are not only compact and efficient but also secure and resilient to adversarial challenges. The implementation details and essential instructions are available on DD-RobustBench.

Avoiding lattice oxygen involvement (oxygen redox) while promoting the coupling of adjacent adsorbed oxygen (metal redox) during the acidic oxygen evolution reaction (OER) is essential for gaining high activity and robust stability in RuO2‐based catalysts but remains elusive. Here, we present a precise strategy to selectively activate the metal redox process while suppressing the undesired oxygen redox pathway by fine‐tuning the Ru–O coordination number in amorphous RuOx. The optimized catalyst exhibits outstanding acidic OER performance, achieving a low overpotential of 215 mV at 10 mA cm⁻² and maintaining stability for 300 h with a negligible degradation rate of 100 µV h⁻¹. X‐ray absorption measurements and multiple operando spectra reveal that only Ru2–O11 moieties can selectively activate the metal redox process, whereas Ru2–O9 and Ru2–O8 moieties either trigger both redox pathways or bypass them. Theoretical calculations reveal that Ru2–O11 moiety reduces crystal field splitting energy at active Ru sites, disables lattice oxygen activation, and lowers the energy barrier for oxygen coupling. The strategy developed in this work offers new avenues for switching redox centers and refining OER mechanisms to enhance catalytic performance and long‐term stability.

Metal catalysts for the CO2 reduction reaction (CO2RR) face challenges such as high cost, limited durability, and environmental impact. Although various structurally diverse and functional metal‐free catalysts have been developed, they often suffer from slow kinetics, low selectivity, and nonrecyclability, significantly limiting their practical applications. In this study, we introduce a recyclable nonmetallic polymer material (vitrimer) as a catalyst for a new platform in contact‐electrocatalysis. This approach harnesses the contact charges generated between water droplets and vitrimer to drive CO2RR, achieving methanol selectivity exceeding 90%. The imine groups within the vitrimer play a dual role, facilitating CO2 adsorption and enriching friction‐generated electrons, thereby mediating efficient electron transfer between the imine groups and CO2 to promote CO2RR. After 84 h of CO2RR, the system achieved a methanol production rate of 13 nmol·h⁻¹, demonstrating the excellent stability of the method. Moreover, the vitrimer retains its high‐performance electrocatalytic activity even after recycling. Mechanistic studies reveal that, compared to traditional metal catalysts, the N─O bond in the imine, which adsorbs the key intermediate *OCH3, breaks more readily to produce methanol, resulting in enhanced product selectivity and yield. This efficient and environmentally friendly contact‐electroreduction strategy for CO2 offers a promising pathway toward a circular carbon economy by leveraging natural water droplet‐based contact‐electrochemistry.

The early and accurate detection of breast lesions through mammography is crucial for improving survival rates. However, the existing deep learning-based methods often rely on costly pixel-level annotations, limiting their scalability in real-world applications. To address this issue, a novel local extremum mapping (LEM) mechanism is proposed for mammogram classification and weakly supervised lesion localization. The proposed method first divides the input mammogram into multiple regions and generates score maps through convolutional neural networks. Then, it identifies the most informative regions by filtering local extrema in the score maps and aggregating their scores for final classification. This strategy enables lesion localization with only image-level labels, significantly reducing annotation costs. Experiments on two public mammography datasets, CBIS-DDSM and INbreast, demonstrate that the proposed method achieves competitive performance. On the INbreast dataset, LEM improves classification accuracy to 96.3% with an AUC of 0.976. Furthermore, the proposed method effectively localizes lesions with a dice similarity coefficient of 0.37, outperforming Grad-CAM and other baseline approaches. These results highlight the practical significance and potential clinical applications of our approach, making automated mammogram analysis more accessible and efficient.

Electrocatalytic CO reduction (COR) offers a promising alternative approach for synthesizing valuable chemicals, potentially at a lower carbon intensity as compared to conventional chemical production. Cu‐based catalysts have shown encouraging selectivity and activity toward multi‐carbon (C2+) products, albeit typically in the form of a mixture. Steering COR selectivity toward specific types of C2+ products, such as liquid products with high energy density, remains a challenge. In this study, we developed a Cu/Zn bimetallic catalyst composite and demonstrated enhanced selectivity toward liquid products as compared to reference CuO and Cu‐based catalysts, approaching 60% at a high current density of 300 mA/cm². Our investigation highlights that the introduction of Zn promoted the emergence of a Cu/Zn heterojunction interface during COR. Density functional theory simulations were used to rationalize the observed differences in selectivity, revealing that interface plays a crucial role in diminishing the oxygen adsorption at the Cu‐sites and modifying the adsorption energy of COR reaction intermediates, consequently leading to enhanced selectivity toward liquid products.

We report the first experimental detection of a zero-frequency fluctuation that is pumped by an Alfvén mode in a magnetically confined plasma. Core-localized Alfvén modes of frequency inside the toroidicity-induced gap (and its harmonics) exhibit three-wave coupling interactions with a zero-frequency fluctuation. The observation of the zero-frequency fluctuation is consistent with theoretical and numerical predictions of zonal modes pumped by Alfvén modes, and is correlated with an increase in the deep core ion temperature, temperature gradient, confinement factor H 89 , P , and a reduction in the main ion heat diffusivity. Despite the energetic particle transport induced by the Alfvén eigenmodes, the generation of a zero-frequency fluctuation that can suppress the turbulence leads to an overall improvement of confinement. Published by the American Physical Society 2025

As the scenarios for electrocardiogram (ECG) monitoring become increasingly diverse, particularly with the development of wearable ECG, the influence of ambiguous factors in diagnosis has been amplified. Reliable ECG information must be extracted from abundant noises and confusing artifacts. To address this issue, we suggest an uncertainty-inspired model for beat-level diagnosis (UI-Beat). The base architecture of UI-Beat separates heartbeat localization and event diagnosis in two branches to address the problem of heterogeneous data sources. To disentangle the epistemic and aleatoric uncertainty within one stage in a deterministic neural network, we propose a new method derived from uncertainty formulation and realize it by introducing the class-biased transformation. Then the disentangled uncertainty can be utilized to screen out noise and identify ambiguous heartbeat synchronously. The results indicate that UI-Beat can significantly improve the performance of noise detection (from 91.60% to 97.50% for real-world noise detection and from 61.40% to 82.41% for real-world artifact detection). For multi-lead ECG analysis, UI-Beat is approaching the performance upper bound in heartbeat localization (only 15 false positives and 9 false negatives out of the 175,907 heartbeats in the INCART database) and achieving a significant performance improvement in heartbeat classification through uncertainty-based cross-lead fusion compared to single-lead prediction and other state-of-the-art methods (an average improvement of 14.28% for detecting heartbeats of S and 3.37% for detecting heartbeats of V). Considering the characteristic of one-stage ECG analysis within one model, it is suggested that the proposed UI-Beat has the potential to be employed as a general model for arbitrary scenarios of ECG monitoring, with the capacity to remove invalid episodes, and realize heartbeat-level diagnosis with confidence provided.