Prior Domain Knowledge Research Articles

DecTree: a physics-based geochemical surrogate for surface complexation of uranium on clay

Abstract. Geochemistry is usually the computational bottleneck in coupled reactive transport simulations, which hampers the complexity of the systems and of the processes they can investigate. In recent years, promising speedups have been obtained by substituting the numerical solution of geochemical models with approximated surrogates borrowed from artificial intelligence and machine learning (AI/ML). In the framework of the DONUT/EURAD project a set of benchmarks were defined to assess the performance and the accuracy of different surrogate approaches in settings relevant to the safety assessment of nuclear waste repositories, such as the surface complexation and exchange of U(VI) on clay. In this context, this work introduces am original surrogate modelling approach based on recursive partitioning of parameter space, which exploits prior domain knowledge for the training. The surrogate, which can be represented as a decision tree, hence the DecTree name, performs dimensionality reduction by identifying functional relationships between outputs and input variables using a straightforward non-monotonic extension of the Spearman's rank correlation coefficient. New predictions are then interpolated from the partitioned training data. Applied to a low-dimensional geochemical model, DecTree shows virtually no training time and excellent accuracy, ensuring a throughput of around 500 000 predictions per second on a single CPU core.

MeDi‐TODER: Medical Domain‐Incremental Task‐Oriented Dialogue Generator Using Experience Replay

ABSTRACTArtificial intelligence (AI) technology has brought groundbreaking changes to the healthcare domain. Specifically, the integration of a medical dialogue system (MDS) has facilitated interactions with patients, identifying meaningful information such as symptoms and medications from their dialogue history to generate appropriate responses. However, shortcomings arise when MDS lacks access to the patient's cumulative history or prior domain knowledge, resulting in the generation of inaccurate responses. To address this challenge, we propose a medical domain‐incremental task‐oriented dialogue generator using experience replay (MeDi‐TODER) that applies the continual learning technique to the medical task‐oriented dialogue generator. By strategically sampling and storing exemplars from previous domains and rehearsing it as it learns, the model effectively retains knowledge and can respond to the novel domains. Extensive experiments demonstrated that MeDi‐TODER significantly outperforms other models that lack continual learning in both natural language generation and natural language understanding. Notably, our proposed method achieves the highest scores, surpassing the upper‐class benchmarks.

Physics-Informed Neural Networks-Based Online Excavation Trajectory Planning for Unmanned Excavator

As a large-scale mining excavator, the electric shovel (ES) has been extensively employed in open-pit mines for overburden removal and mineral loading. In the development of unmanned operations for ES, dynamic excavation trajectory planning is essential, as it directly influences operational efficiency and energy consumption by guiding the dipper during excavation. However, conventional optimization-based methods for excavation trajectory planning typically start from scratch, resulting in a time-consuming process that fails to meet real-time requirements. To address this challenge, we propose an innovative online trajectory planning framework based on physics-informed neural networks (PINNOTP) that utilizes advanced data-driven techniques. The input to PINNOTP consists of on-site working conditions, including the initial state of the ES and the material surface being excavated. The output is a smooth, polynomial-based curve that serves as the reference trajectory for the dipper. To ensure smooth execution of the generated trajectory, prior domain knowledge—such as physics-based target-oriented constraints, essential system dynamics, and mechanical constraints—is explicitly incorporated into the loss function during training. A case study is presented to validate the proposed method, demonstrating that PINNOTP effectively addresses the challenges of online excavation trajectory planning.

Prior knowledge and new learning: An experimental study of domain-specific knowledge.

It is commonly claimed that higher domain knowledge enhances new learning-the knowledge-is-power hypothesis. However, a recent meta-analysis (Simonsmeier et al., 2022) has challenged this idea, finding no overall relationship between prior knowledge and new learning across hundreds of highly variable effect sizes. The authors note that this variability and lack of randomized controlled experiments preclude broad claims regarding the influence of prior knowledge on learning. The present study (conducted in 2020) provides an experimental assessment of the causal effect of prior domain knowledge on new learning. Participants were randomly assigned to receive training in one of two academic domains over 3 days before learning new information about topics in both domains for a later test. Training was specific to three of four topics within that domain, allowing the untrained topic in the trained domain to act as a measure of new learning in that domain. New learning, measured as final test performance or knowledge gains, did not differ between the high and low domain knowledge conditions. Experimentally induced prior domain knowledge did not affect new learning. (PsycInfo Database Record (c) 2024 APA, all rights reserved).

Feature-level Fusion vs. Score-level Fusion for Image Retrieval Based on Pre-trained Deep Neural Networks

Today’s complex multimedia content made retrieving images similar to the user’s query from the database a challenging task. The performance of a Content-Based Image Retrieval System (CBIR) system highly depends on the image representation in a form of low-level features and similarity measurement. The traditional visual descriptors that do not provide good prior domain knowledge could lead to poor performance retrieval results. On the other hand, Deep Convolutional Neural Networks (DCNNs) have recently achieved a remarkable success as methods for image classification in various domains. Recently, pre-trained deep convolution neural networks on thousands of classes have the ability to extract very accurate and representative features which, in addition to classification, can also be successfully used in image retrieval systems. ResNet152, GoogLeNet and InceptionV3 are some of the effective and successful examples of pre-trained DCNNs recently applied in a computer vision tasks such as object recognition, clustering, and classification. In this paper, two approaches for a CBIR system, namely early fusion and late fusion, have been presented and compared. The early fusion utilizes concatenation of the features extracted by each possible pair of DCNNs, that is ResNet152-GoogLeNet, ResNet152-InceptionV3, and GoogLeNet-InceptionV3, and the late fusion apply CombSum method with Z-Score standardization to combine the score results provided by each DCNN of the aforementioned pairs. In the experiments on a popular WANG dataset it has been shown that late fusion approach slightly outperforms early fusion approach. The best performance of our experiments in terms of Average Precision (AP) for the top 20 results reaches 96.82%.

Enhancing chemistry-intuitive feature learning to improve prediction performance of optical properties.

Emitters have been widely applied in versatile fields, dependent on their optical properties. Thus, it is of great importance to explore a quick and accurate prediction method for optical properties. To this end, we have developed a state-of-the-art deep learning (DL) framework by enhancing chemistry-intuitive subgraph and edge learning and coupling this with prior domain knowledge for a classic message passing neural network (MPNN) which can better capture the structural features associated with the optical properties from a limited dataset. Benefiting from technical advantages, our model significantly outperforms eight competitive ML models used in five different optical datasets, achieving the highest accuracy to date in predicting four important optical properties (absorption wavelength, emission wavelength, photoluminescence quantum yield and full width at half-maximum), showcasing its robustness and generalization. More importantly, based on our predicted results, one new deep-blue light-emitting molecule PPI-2TPA was successfully synthesized and characterized, which exhibits close consistency with our predictions, clearly confirming the application potential of our model as a quick and reliable prediction tool for the optical properties of diverse emitters in practice.

Analogy use in textbook introductions to isostasy: Presentations, perceptions and physics

In teaching, analogies can be used to support the introduction of a new concept or idea with reference to a more familiar one. This relies on the learner’s correct understanding of the (supposedly) familiar situation, their identification of the relevant similarities (structural relations) between the two situations, and of the limits of the similarities. Without these, analogies may reinforce or introduce misconceptions. Through content analysis of 14 introductory textbooks, semi-structured interviews with four students, and analysis of two analogies, we explore how analogies are used as part of introductions to isostasy in introductory geoscience textbooks. Analogies were ubiquitous, but prior knowledge of the source domain was typically assumed and the majority were not elaborated (explained to the learner). Interviews indicate that, in the absence of sufficient elaboration, students turn to superficial similarities for clues on how to apply the analogies. Many textbooks used analogies amenable to this approach, indicating an alignment between textbook strategy and student reactions. However, giving precedence to superficial similarities over explicit elaboration of structural relations may favor a novice approach to learning. Analyzing an analogy using free-body diagrams demonstrates how an apparently relevant analogy can introduce misconceptions, both in terms of the underlying physics (introducing a fictitious force) and the Earth application (unnatural behavior of the lithosphere) when not sufficiently, or correctly, elaborated. We conclude that the introductory textbooks place much of the responsibility for appropriate use of their analogies, and for avoiding misconceptions, on the learners and their teachers.

People also ask: How does this tool affect exploration-exploitation strategies with regard to prior domain knowledge and search context? An eye-tracking study

With the diversification of Internet uses, online content type has become richer. Alongside organic results, search engine results pages now provide tools to improve information searching and learning. The People also ask (PAA) box is intended to reduce users' cognitive costs by offering easily accessible information. Nevertheless, there has been scant research on how users actually process it, compared with more traditional content type (i.e., organic results and online documents). The present eye-tracking study explored this question by considering the search context (complex lookup task vs. exploratory task) and users' prior domain knowledge (high vs. low). Main results show that users fixated the PAA box and online documents more to achieve exploratory goals, and fixated organic results more to achieve lookup goals. Users with low knowledge process PAA content at an early stage in their search contrary to their counterparts with high knowledge. Given these results, information system developers should diversify PAA content according to search context and users’ prior domain knowledge.

A Genetic Causal Explainer for Deep Knowledge Tracing

Knowledge tracing (KT) has become an increasingly relevant problem in intelligent education services. Deep learning-based knowledge tracing (DLKT) achieves superb performance in terms of prediction accuracy, but it lacks of explainability, which makes us hard to trust or understand models. The previous work on explaining DLKT was mainly based on gradients or attention scores, which is susceptible to spurious correlations, reducing the credibility of the explanation. To address this limitation, in this paper, we propose a causal explanation method based on the genetic algorithm (GA), named Genetic Causal Explainer(GCE), which constructs a causal framework to estimate the attribution of subsequence to the predictions of DLKT models, and a genetic coding system is designed. Further, A multi-strategy initialization method inspired by domain prior knowledge is proposed, and a global empirical matrix is introduced to capture the causal correlation knowledge during the search process across instances, and guiding the mutation operators. The GCE as a post hoc explanation method can generate explanation results without affecting model training, and can be applied to analyze different DLKT models. Experimental results demonstrate the GCE perform better than other explanation methods in terms of accuracy and readability in quantitative assessments. Meanwhile, the GCE also shows good application prospects in mining educational laws and comparing KT models.

Modeling 0.6 million genes for the rational design of functional cis-regulatory variants and de novo design of cis-regulatory sequences

Rational design of plant cis-regulatory DNA sequences without expert intervention or prior domain knowledge is still a daunting task. Here, we developed PhytoExpr, a deep learning framework capable of predicting both mRNA abundance and plant species using the proximal regulatory sequence as the sole input. PhytoExpr was trained over 17 species representative of major clades of the plant kingdom to enhance its generalizability. Via input perturbation, quantitative functional annotation of the input sequence was achieved at single-nucleotide resolution, revealing an abundance of predicted high-impact nucleotides in conserved noncoding sequences and transcription factor binding sites. Evaluation of maize HapMap3 single-nucleotide polymorphisms (SNPs) by PhytoExpr demonstrates an enrichment of predicted high-impact SNPs in cis-eQTL. Additionally, we provided two algorithms that harnessed the power of PhytoExpr in designing functional cis-regulatory variants, and de novo creation of species-specific cis-regulatory sequences through in silico evolution of random DNA sequences. Our model represents a general and robust approach for functional variant discovery in population genetics and rational design of regulatory sequences for genome editing and synthetic biology.

NPEX: Never give up protein exploration with deep reinforcement learning

Elucidating unknown structures of proteins, such as metastable states, is critical in designing therapeutic agents. Protein structure exploration has been performed using advanced computational methods, especially molecular dynamics and Markov chain Monte Carlo simulations, which require untenably long calculation times and prior structural knowledge. Here, we developed an innovative method for protein structure determination called never give up protein exploration (NPEX) with deep reinforcement learning. The NPEX method leverages the soft actor-critic algorithm and the intrinsic reward system, effectively adding a bias potential without the need for prior knowledge. To demonstrate the method's effectiveness, we applied it to four models: a double well, a triple well, the alanine dipeptide, and the tryptophan cage. Compared with Markov chain Monte Carlo simulations, NPEX had markedly greater sampling efficiency. The significantly enhanced computational efficiency and lack of prior domain knowledge requirements of the NPEX method will revolutionize protein structure exploration.

Biologically informed NeuralODEs for genome-wide regulatory dynamics

BackgroundGene regulatory network (GRN) models that are formulated as ordinary differential equations (ODEs) can accurately explain temporal gene expression patterns and promise to yield new insights into important cellular processes, disease progression, and intervention design. Learning such gene regulatory ODEs is challenging, since we want to predict the evolution of gene expression in a way that accurately encodes the underlying GRN governing the dynamics and the nonlinear functional relationships between genes. Most widely used ODE estimation methods either impose too many parametric restrictions or are not guided by meaningful biological insights, both of which impede either scalability, explainability, or both.ResultsWe developed PHOENIX, a modeling framework based on neural ordinary differential equations (NeuralODEs) and Hill-Langmuir kinetics, that overcomes limitations of other methods by flexibly incorporating prior domain knowledge and biological constraints to promote sparse, biologically interpretable representations of GRN ODEs. We tested the accuracy of PHOENIX in a series of in silico experiments, benchmarking it against several currently used tools. We demonstrated PHOENIX’s flexibility by modeling regulation of oscillating expression profiles obtained from synchronized yeast cells. We also assessed the scalability of PHOENIX by modeling genome-scale GRNs for breast cancer samples ordered in pseudotime and for B cells treated with Rituximab.ConclusionsPHOENIX uses a combination of user-defined prior knowledge and functional forms from systems biology to encode biological “first principles” as soft constraints on the GRN allowing us to predict subsequent gene expression patterns in a biologically explainable manner.

Domain base dynamic convolution and distance map guidance for anterior mediastinal lesion segmentation

The automatic segmentation of anterior mediastinal lesions in the enhanced CT image plays a significant role in clinical diagnosis. Due to the low incidence rate of anterior mediastinal disease, the data for training segmentation network may come from different hospitals, different equipment, and different operators, which leads to lower generalization and robustness of the trained network. This paper proposes an anterior mediastinal lesion segmentation method based on domain-adaptive dynamic convolution and distance map guidance. Considering the diversity of clinical data sources and equipment differences, the domain prior knowledge is added to the global average-pooling layer of the network, and the dynamic convolution forms specific network parameters for specific domains to enhance domain-specific information. Meanwhile, the Singed Distance Map (SDM) head is involved in the network according to the shape, size and position of the anterior mediastinal lesions to constraint the boundary and assist in localization. Ablation experiments demonstrate that the distance map can effectively constrain the segmentation target and reduce false positives. Both qualitative and quantitative experimental results indicate that our method can achieve more accurate anterior mediastinal lesion segmentation with greater generalization ability. Our approach achieves an overall Dice coefficient of 88.33%, which is 2.31% higher than existing state-of-the-art methods, and has achieved good performance in terms of ASSD evaluation for segmentation edges.

A survey of deep learning-driven architecture for predictive maintenance

Over the past decades, deep learning techniques have attracted increased attention from various research and industrial domains aligned with the development of Industry Internet-of-Things(IIoT). Specifically, with the advantage of data-driven methods, industrial organizations are seeking novel proactive strategies supported by analytic models to guarantee the quality of their production by observing degradation or predicting failure ahead of the occurrence of the component or asset. Predictive strategies are expected to promise the influence of unnecessary maintenance interruptions and mitigate the consequence of that, hence, extending the remaining useful life of products. This paper conducts a survey of the utilization of deep learning technologies on engineering applications where they provide satisfactory solutions with respect to specific data types or input signals. 106 primary papers are reviewed on deep learning–driven approaches which mainly explore five of the most popular architectures in the application of predictive maintenance. The main content of this paper summarizes the common advantages of each architecture and, accordingly, points out their limitations, as well as describes the application scopes of fully connected deep neural networks, convolutional neural networks, stacked autoencoders, deep belief networks, and deep recurrent neural networks. Based on the technique discussion for each of them, we intend to provide a comprehensive understanding and guidance of the appropriate usage of deep learning architectures to devise an effective predictive maintenance strategy for the scientific and industrial developers whose expertise lies in the prior domain knowledge of multi-source isomerization data. Moreover, the main content demonstrated the summarization of the decisive factor by which the incremental stages of the approaches were determined, fundamentally including the dataset specification, feature extraction, and the integration of deep learning approaches.

Adaptive threshold optimisation for online feature selection using dynamic particle swarm optimisation in determining feature relevancy and redundancy

In the era of data-driven decision-making, managing dynamic data streams characterised by evolving data distributions and high dimensionality presents a formidable challenge for online feature selection. This research addresses the challenge by developing innovative solutions in optimising Online Feature Selection (OFS) to manage feature irrelevancy and redundancy and rigorously validating the proposed method in high-dimensional dynamic data streams. The research employs a structured methodology, introducing a novel method: a Dynamic Particle Swarm Optimisation (PSO)-based Threshold Optimisation method. Dynamic PSO, injected on three benchmark methods of OFS (i.e., Online Streaming Feature Selection (OSFS), Fast-OSFS, and Scalable and Accurate Online Feature Selection for Big Data (SAOLA)), enabled the global best position to be dynamically adjusted to exploit the best solution found in streaming data. This research presents two optimisation variants of the proposed method: RedundantPSO, which optimises redundant features, and IrrelevantPSO, which optimises irrelevant features. Unlike the traditional PSO method on feature selection that uses feature encoding representation, the proposed method is underpinned by two contributions: adaptive threshold particle representation of particle swarm optimisation and enhanced fitness function using minimisation of mean absolute deviation of dependency among feature subsets. Adaptive threshold particle representation combines the feature encoding part with a novel aspect that defines a threshold value of significance level ranging from 0.01 to 0.1. This unique contribution sets the research apart in the field where it enables the adaptive adjustment of the threshold value based on incoming features. Next, the adaptation of Mean Absolute Deviation (MAD) was integrated into the fitness evaluation of PSO to gain a more accurate and reliable measure of fitness for feature selection. During the experiment phase, we analysed various benchmark datasets with highly redundant and relevant behaviour. Our analysis concluded that selecting the appropriate threshold values significantly improved model performance for high-redundancy datasets, highlighting the need for careful threshold selection. The experimental evaluations have revealed that integrating RedundantPSO with OSFS (OSFS+RedundantPSO) resulted in a remarkable enhancement of the OSFS method's accuracy, achieving an impressive average accuracy rate of 76.8%. This substantial improvement includes occasional spikes of up to 3.8% over the baseline OSFS accuracy, showcasing OSFS+RedundantPSO as the top-performing combination. Furthermore, Fast-OSFS + RedundantPSO outperformed Fast-OSFS by a slight margin, reaching an average accuracy of 72.7%, while SAOLA + RedundantPSO exhibited a substantial 3.1% increase in average accuracy over SAOLA, reaching 74.0%. It is noteworthy to highlight that the threshold value searched in the proposed method also significantly impacted the identification of the behaviour of the dataset, either high relevancy or redundancy, even in the absence of prior domain knowledge. A higher threshold signifies the evaluation of a more redundant feature space. In conclusion, the results demonstrated the significant contributions of the method in enhancing model accuracy, adapting to evolving data distributions, and optimising feature subsets with acceptable runtime. The research aims to advance the field of data science, such as cybersecurity, finance, healthcare and more, while empowering end-users to make informed decisions under changing data stream circumstances.

Two-path target-aware contrastive regression for action quality assessment

Action quality assessment (AQA) is a challenging vision task due to the complexity and variance of the scoring rules embedded in the videos. Recent approaches have reduced the prediction difficulty of AQA via learning action differences between videos, but there are still challenges in learning scoring rules and capturing feature differences. To address these challenges, we propose a two-path target-aware contrastive regression (T2CR) framework. We propose to fuse direct and contrastive regression and exploit the consistency of information across multiple visual fields. Specifically, we first directly learn the relational mapping between global video features and scoring rules, which builds occupational domain prior knowledge to better capture local differences between videos. Then, we acquire the auxiliary visual fields of the videos through sparse sampling to learn the commonality of feature representations in multiple visual fields and eliminate the effect of subjective noise from a single visual field. To demonstrate the effectiveness of T2CR, we conduct extensive experiments on four AQA datasets (MTL-AQA, FineDiving, AQA-7, JIGSAWS). Our method is superior to state-of-the-art methods without elaborate structural design and fine-grained information.

Data-Driven Overlapping-Track Profile Modeling in Cold Spray Additive Manufacturing

Cold spray additive manufacturing is an emerging solid-state deposition process that enables large-scale components to be manufactured at high-production rates. Control over geometry is important for reducing the development and growth of defects during the 3D build process and improving the final dimensional accuracy and quality of components. To this end, a machine learning approach has recently gained interest in modeling additively manufactured geometry; however, such a data-driven modeling framework lacks the explicit consideration of a depositing surface and domain knowledge in cold spray additive manufacturing. Therefore, this study presents surface-aware data-driven modeling of an overlapping-track profile using a Gaussian Process Regression model. The proposed Gaussian Process modeling framework explicitly incorporated two relevant geometric features (i.e., surface type and polar length from the nozzle exit to the surface) and a widely adopted Gaussian superposing model as prior domain knowledge in the form of an explicit mean function. It was shown that the proposed model could provide better predictive performance than the Gaussian superposing model alone and the purely data-driven Gaussian Process model, providing consistent overlapping-track profile predictions at all overlapping ratios. By combining accurate prediction of track geometry with toolpath planning, it is anticipated that improved geometric control and product quality can be achieved in cold spray additive manufacturing.

Physics-data coupling-driven method to predict the penetration depth into concrete targets

The projectile penetration process into concrete target is a nonlinear complex problem. With the increase of experiment data, the data-driven paradigm has exhibited a new feasible method to solve such complex problem. However, due to poor quality of experimental data, the traditional machine learning (ML) methods, which are driven only by experimental data, have poor generalization capabilities and limited prediction accuracy. Therefore, this study intends to exhibit a ML method fusing the prior knowledge with experiment data. The new ML method can constrain the fitting to experimental data, improve the generalization ability and the prediction accuracy. Experimental results show that integrating domain prior knowledge can effectively improve the performance of the prediction model for penetration depth into concrete targets.

Automatic recognition of white blood cell images with memory efficient superpixel metric GNN: SMGNN.

An automatic recognizing system of white blood cells can assist hematologists in the diagnosis of many diseases, where accuracy and efficiency are paramount for computer-based systems. In this paper, we presented a new image processing system to recognize the five types of white blood cells in peripheral blood with marked improvement in efficiency when juxtaposed against mainstream methods. The prevailing deep learning segmentation solutions often utilize millions of parameters to extract high-level image features and neglect the incorporation of prior domain knowledge, which consequently consumes substantial computational resources and increases the risk of overfitting, especially when limited medical image samples are available for training. To address these challenges, we proposed a novel memory-efficient strategy that exploits graph structures derived from the images. Specifically, we introduced a lightweight superpixel-based graph neural network (GNN) and broke new ground by introducing superpixel metric learning to segment nucleus and cytoplasm. Remarkably, our proposed segmentation model superpixel metric graph neural network (SMGNN) achieved state of the art segmentation performance while utilizing at most 10000$ \times $ less than the parameters compared to existing approaches. The subsequent segmentation-based cell type classification processes showed satisfactory results that such automatic recognizing algorithms are accurate and efficient to execeute in hematological laboratories. Our code is publicly available at https://github.com/jyh6681/SPXL-GNN.

Prior Knowledge-guided Triple-Domain Transformer-GAN for Direct PET Reconstruction from Low-Count Sinograms.

To obtain high-quality positron emission tomography (PET) images while minimizing radiation exposure, numerous methods have been dedicated to acquiring standard-count PET (SPET) from low-count PET (LPET). However, current methods have failed to take full advantage of the different emphasized information from multiple domains, i.e., the sinogram, image, and frequency domains, resulting in the loss of crucial details. Meanwhile, they overlook the unique inner-structure of the sinograms, thereby failing to fully capture its structural characteristics and relationships. To alleviate these problems, in this paper, we proposed a prior knowledge-guided transformer-GAN that unites triple domains of sinogram, image, and frequency to directly reconstruct SPET images from LPET sinograms, namely PK-TriDo. Our PK-TriDo consists of a Sinogram Inner-Structure-based Denoising Transformer (SISD-Former) to denoise the input LPET sinogram, a Frequency-adapted Image Reconstruction Transformer (FaIR-Former) to reconstruct high-quality SPET images from the denoised sinograms guided by the image domain prior knowledge, and an Adversarial Network (AdvNet) to further enhance the reconstruction quality via adversarial training. Specifically tailored for the PET imaging mechanism, we injected a sinogram embedding module that partitions the sinograms by rows and columns to obtain 1D sequences of angles and distances to faithfully preserve the inner-structure of the sinograms. Moreover, to mitigate high-frequency distortions and enhance reconstruction details, we integrated global-local frequency parsers (GLFPs) into FaIR-Former to calibrate the distributions and proportions of different frequency bands, thus compelling the network to preserve high-frequency details. Evaluations on three datasets with different dose levels and imaging scenarios demonstrated that our PK-TriDo outperforms the state-of-the-art methods.