Knowledge-Augmented Adaptive Mechanism for Radiology Report Generation

Predicting the risk of potential diseases from Electronic Health Records (EHR) has attracted considerable attention in recent years, especially with the development of deep learning techniques. Compared with traditional machine learning models, deep learning based approaches achieve superior performance on risk prediction task. However, none of existing work explicitly takes prior medical knowledge (such as the relationships between diseases and corresponding risk factors) into account. In medical domain, knowledge is usually represented by discrete and arbitrary rules. Thus, how to integrate such medical rules into existing risk prediction models to improve the performance is a challenge. To tackle this challenge, we propose a novel and general framework called PRIME for risk prediction task, which can successfully incorporate discrete prior medical knowledge into all of the state-of-the-art predictive models using posterior regularization technique. Different from traditional posterior regularization, we do not need to manually set a bound for each piece of prior medical knowledge when modeling desired distribution of the target disease on patients. Moreover, the proposed PRIME can automatically learn the importance of different prior knowledge with a log-linear model.Experimental results on three real medical datasets demonstrate the effectiveness of the proposed framework for the task of risk prediction

The separation of a lung tumor from adjacent normal tissue, which has similar intensity values and indistinct boundaries on low-contrast CT images is a challenging task. In this paper, a prior knowledge enhanced random walk (RW) is proposed to account for the prior functional knowledge from PET and intensity information from CT. The prior knowledge acquired from PET is used for the automated selection of foreground seeds, defined as the tumor confidence region, the background seeds and the walking range to increase computational efficiency of the RW algorithm in CT. The tumor confidence region is also used for balancing transition, and thus limiting the information propagation range through a weight factor. The experimental evaluation on 18 low-contrast CT images with manual tumor segmentation demonstrated that our method outperformed RW and random walk from restart (RWR) as measured by the Dice similarity coefficient (DSC).

BackgroundArterial calcification is associated with cardiovascular disease as a complication of advanced atherosclerosis and is a significant contributor to cardiovascular morbidity and mortality. Osteoblastic differentiation of vascular smooth muscle cells (VSMCs) plays an important role in arterial calcification and is characterized by cellular necrosis, inflammation, and lipoprotein and phospholipid complexes, especially in atherosclerotic calcification. The conditioned medium from bone marrow-derived mesenchymal stem cells (MSC-CM) is well known as a rich source of autologous cytokines and is universally used for tissue regeneration in current clinical medicine. Here, we demonstrate that MSC-CM inhibits beta-glycerophosphate (β-GP)-induced vascular calcification through blockade of the bone morphogenetic protein-2 (BMP2)–Smad1/5/8 signaling pathway.MethodsVSMC calcification was induced by β-GP followed by treatment with MSC-CM. Mineral deposition was assessed by Alizarin Red S staining. Intracellular calcium content was determined colorimetrically by the o-cresolphthalein complexone method and alkaline phosphatase (ALP) activity was measured by the para-nitrophenyl phosphate method. Expression of BMP2, BMPR1A, BMPR1B, BMPR2, msh homeobox 2 (Msx2), Runt-related transcription factor 2 (Runx2), and osteocalcin (OC), representative osteoblastic markers, was assessed using real-time polymerase chain reaction analysis while the protein expression of BMP2, Runx2, and phosphorylated Smad1/5/8 was detected by western blot analysis.ResultsOur data demonstrated that MSC-CM inhibits osteoblastic differentiation and mineralization of VSMCs as evidenced by decreased calcium content, ALP activity, and decreased expression of BMP-2, Runx2, Msx2, and OC. MSC-CM suppressed the expression of phosphorylated Smad1/5/8 and the β-GP-induced translocation from the cytoplasm to the nucleus. Further study demonstrated that human recombinant BMP-2 overcame the suppression of VSMC calcification by MSC-CM.ConclusionMSC-CM may act as a novel therapy for VSMC calcification by mediating the BMP2–Smad1/5/8 signaling pathway

The platelet-rich fibrin (PRF) is known as a rich source of autologous cytokines and growth factors and universally used for tissue regeneration in current clinical medicine. However, the microstructure of PRF has not been fully investigated nor have been studied the key molecules that differ PRF from platelet-rich plasma. We fabricated PRF under Choukroun's protocol and produced its extract (PRFe) by freezing at -80°C. The conventional histological, immunohistological staining, and scanning electron microscopy images showed the microstructure of PRF, appearing as two zones, the zone of platelets and the zone of fibrin, which resembled a mesh containing blood cells. The PRFe increased proliferation, migration, and promoted differentiation of the human alveolar bone marrow stem cells (hABMSCs) at 0.5% concentration in vitro. From the results of proteome array, matrix metalloproteinase 9 (MMP9) and Serpin E1 were detected especially in PRFe but not in concentrated platelet-rich plasma. Simultaneous elevation of MMP9, CD44, and transforming growth factor β-1 receptor was shown at 0.5% PRFe treatment to the hABMSC in immunoblot. Mineralization assay showed that MMP9 directly regulated mineralization differentiation of hABMSC. Transplantation of the fresh PRF into the mouse calvarias enhanced regeneration of the critical-sized defect. Our results strongly support the new characteristics of PRF as a bioscaffold and reservoir of growth factors for tissue regeneration.

This dissertation explores the intersection of data fusion, federated learning, and Bayesian methods, with a focus on their applications in indoor localization, GNSS, and image processing. Data fusion involves integrating data and knowledge from multiple sources. It becomes essential when data is only available in a distributed fashion or when different sensors are used to infer a quantity of interest. Data fusion typically includes raw data fusion, feature fusion, and decision fusion. In this thesis, we will concentrate on feature fusion. Distributed data fusion involves merging sensor data from different sources to estimate an unknown process. Bayesian framework is often used because it can provide an optimal and explainable feature by preserving the full distribution of the unknown given the data, called posterior, over the estimated process at each agent. This allows for easy and recursive merging of sensor data with prior knowledge without the need for storage. However, there are challenges in data fusion, such as the multiple counting problem, which occurs when data is utilized numerous times without the user's knowledge. In a Bayesian setting, a priori information of the unknown quantities is available and can be present among different distributed estimators. When the local estimates are fused, the prior knowledge used to construct several local posteriors might be overused unless the fusion node accounts for this and corrects it. In this thesis, we analyze the effects of shared priors in Bayesian data fusion contexts. Our analysis helps to understand the performance behavior as a function of the number of collaborative agents and different types of priors, depending on different common fusion rules. To perform this analysis, we use two divergences that are common in Bayesian inference. The generality of the results allows us to analyze very generic distributions. These theoretical results are corroborated through experiments in a variety of estimation and classification problems, including linear and nonlinear models, and federated learning schemes in a Bayesian perspective. Federated Learning is a discipline of Distributed Data Fusion that can be used for model training on local data without exchanging the data itself. This technique is useful in scenarios like image classification, indoor positioning, and jammer signal classification, where privacy concerns make data crowdsourcing difficult. In this thesis, we explore FL in two specific applications: Jammer signal and indoor positioning. Jamming signals can disrupt the operation of GNSS receivers, making jamming mitigation and localization techniques essential. Jammer classification can help with these techniques. For indoor positioning, local data points are collected to create a local fingerprinting database. Data-driven machine learning methods are used to predict the location based on this model learned without data sharing. However, Federated Learning faces several challenges, including the non-IID (non-identically distributed) problem. This problem occurs when the data distribution varies across different clients, which can lead to a decrease in performance. In this thesis, we will discuss various FL frameworks and algorithms that can help address this problem. A FL method from Bayesian perspective are proposed to deal with the problem in indoor positioning problem. Additionally, we will delve into the concepts of personalized and clustered Federated Learning. The personalized approach tailors the learning process to individual clients, which enhances the system's performance by leveraging the unique data and characteristics of each client. We will discuss the application of this approach in indoor positioning and jammer signal classification, highlighting its potential improvements and challenges. Expanding on the personalized approach, we introduce and discuss clustered Federated Learning. This learning method groups clients with similar characteristics or data patterns. Each client is typically associated with one data distribution and participates in training a model for that distribution along with their cluster peers. We propose a Bayesian framework for clustered FL, which associates clients to clusters, along with several practical algorithms for efficient learning, considering the trade-offs between performance and computational complexity. In conclusion, this thesis provides a detailed investigation and evaluation of the integration of data fusion, federated learning, and Bayesian methods. The insights and conversations presented in this study are anticipated to enhance the progress and implementation of these technologies in various domains, particularly in indoor positioning, and GNSS.--Author's abstract

The data-driven InfraRed Small Target Detection (IRSTD) methods have witnessed remarkable advancements in performance. However, these methods typically rely on high-quality pixel-level mask labels, requiring substantial human effort and time for annotation. To tackle this challenge, we propose a Prior Knowledge Enhanced Learning Approach (PKELA), which contains two key strategies: the Prior Knowledge Enhanced Initial Pseudo-label Generation (PKEIPG) strategy and the Teacher Knowledge Guided Label Update (TKGLU) strategy. Specifically, for the PKEIPG strategy, we introduce a Local Contrast Enhancement (LCE) module based on prior knowledge to suppress background interference. Drawing inspiration from the human brain’s processing sequence of visual information, we locate an appropriate neighborhood containing the infrared small target by identifying the target’s edges. Within this neighborhood, high-quality initial pseudo labels are obtained through adaptive threshold segmentation, which provide strong supervisory signals for the data-driven IRSTD method during the initial training phase. For the TKGLU strategy, we employ an Exponential Moving Average (EMA) teacher model that dynamically updates its parameters based on the current model’s state. Furthermore, a memory bank is established to archive the teacher model’s performance, which is treated as prior knowledge. This knowledge is then systematically used to guide the refinement of pseudo labels. Experimental results on the SIRST, NUDT-SIRST, IRSTD-1k, and SIRST3 datasets indicate that our PKELA exhibits superior training performance with single-point labels. By incorporating our approach, the data-driven IRSTD methods could achieve near-full-supervised performance in terms of Probability of Detection (<italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">Pd</i>) metric, with Intersection over Union (<italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">IoU</i>) and normalized Intersection over Union (<italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">nIoU</i>) reaching up to 95.75% and 95.82% of full-supervised performance, respectively. Our code is available at https://gitee.com/mynewspace/pkela.

Reinforcement learning has made great success in recent years. Generally, the learning process requires a huge amount of interaction with the environment before an agent can achieve acceptable performance. This motivates many techniques, such as incorporating prior knowledge which is usually presented as experts’ demonstration, and using a probability distribution to represent state-and-action values, to accelerate the learning process. The methods perform well when the prior knowledge is genuinely correct and no much change occurs to the learning environment. However, the requirement is not perfectly realistic in many complex applications. The demonstration knowledge may not reflect the true environment and even be full of noise. In this paper, we introduce a dynamic distribution merging method to improve knowledge utilization in a general reinforcement learning algorithm, namely Q-learning. The new method adapts a normal distribution to represent state-action values and merges the prior and learned knowledge in a discriminative way. We theoretically analyze the new learning method and demonstrate its empirical performance over multiple problem domains.

One of the most central discussions in entrepreneurship concerns the importance of knowledge in opportunity discovery and exploitation. Austrian economics suggests that prior knowledge is related to successful discovery and exploitation. This stream of thought implicitly assumes that opportunity discovery is instantaneous and/or that opportunities remain constant over time. In dynamic markets where change is frequent and sometimes erratic, opportunities become more elusive and rapidly appear, disappear or change. We discuss two longitudinal case studies in order to demonstrate the functioning, integration, and utilization of knowledge in successful and failed opportunity discovery and exploitation. The cases involved are new ventures and we focus on their original business opportunity and their subsequent metamorphoses in the discovery and exploitation processes. Our findings show that while prior knowledge may be of great importance, constantly gaining new market knowledge and adapting strategies in accordance with this new knowledge is essential to the successful discovery and exploitation of opportunities. These findings challenge the common assumption that entrepreneurs first discover opportunities and then exploit them. Instead, entrepreneurial opportunity discovery and exploitation are closely entwined and exploitation influences discovery just as discovery influences exploitation.

K-gap metric is an innovative tool for fault diagnosis system design and analysis. In this paper, different methods of data-driven stable kernel representation (SKR) are investigated for the purpose of K-gap metric calculation. Three existing methods are compared by definitions of SKR and calculation procedures. The first difference between these methods lies in their different definitions of SKR. The comparison result shows that the one which is unified with model-based SKR is the better definition version. The second difference is the different noise reduction techniques applied to process data, which have influences on numerical result. Based on the above comparison, a new method is proposed with more accuracy in K-gap metric calculation by using the prior knowledge from model-based method. Finally, a numerical example compare and illustrate the accuracy in K-gap computation of four data-driven SKR methods.

Geoscience-aware deep learning: A new paradigm for remote sensing

Tomographic imaging is an ill-posed linear inverse problem, and is often regularized using prior knowledge of the sought-after object property. However, typical hand-crafted priors such as sparsity-promoting penalties may be insufficient to comprehensively describe the prior knowledge of the object to-be-imaged. In order to utilize more detailed prior knowledge, data-driven methods using deep neural networks have recently been explored for learning a prior from existing image data. However, an analysis of the ability of such data-driven methods to generalize to data that may lie outside the training distribution is still under investigation. This is particularly critical for medical imaging applications. In order to address such concerns, in this work we propose to understand the effect of the prior imposed by a reconstruction method by comparing the null components of the sought-after object and its reconstructed estimate, when ground truth objects are available. The concept of a hallucination map is introduced for the purpose of assessing non-data-driven and data-driven regularization for image reconstruction. Numerical studies were conducted using stylized undersampled k-space measurements from publicly available magnetic resonance imaging (MRI) datasets. It is demonstrated that the proposed method can be employed to identify the source of false structures in estimates of the sought-after object for a given reconstruction method.

There are problems such as poor interpretability and insufficient generalization ability when extracting buildings from high-resolution remote sensing images based on deep learning. This paper proposes a building extraction model called BPKG-SegFormer (Building Prior Knowledge Guided SegFormer) that combines prior knowledge of buildings with data-driven methods. This model constructs a building feature attention module and utilizes the multi-task loss function to optimize the extraction of buildings. Experimental results show that on the WHU building dataset, the proposed model outperforms UNet, Deeplabv3 + , and SegFormer models with OA, P, R, and MIoU of 96.63%, 95.94%, 94.76%, and 90.6%, respectively. The BPKG-SegFormer model extracts buildings with more regular shapes and flatter edges, reducing internal voids and increasing the number of correctly detected buildings.

Digital twin approach for on-load tap changers using data-driven dynamic model updating and optimization-based operating condition estimation

Robust feature selection to predict tumor treatment outcome.

Background/Objectives: The limited availability of labeled data, particularly in the medical domain, poses a significant challenge for training accurate diagnostic models. While deep learning techniques have demonstrated notable efficacy in image-based tasks, they require large annotated datasets. In data-scarce scenarios—especially involving rare diseases—their performance deteriorates significantly. Meta-learning offers a promising alternative by enabling models to adapt quickly to new tasks using prior knowledge and only a few labeled examples. This study aims to evaluate the effectiveness of representative meta-learning models for thoracic disease classification in chest X-rays. Methods: We conduct a comparative evaluation of four meta-learning models: Prototypical Networks, Relation Networks, MAML, and FoMAML. First, we assess five backbone architectures (ConvNeXt, DenseNet-121, ResNet-50, MobileNetV2, and ViT) using a Prototypical Network. The best-performing backbone is then used across all meta-learning models for fair comparison. Experiments are performed on the ChestX-ray14 dataset under a 2-way setting with multiple k-shot configurations. Results: Prototypical Networks combined with DenseNet-121 achieved the best performance, with a recall of 68.1%, an F1-score of 67.4%, and a precision of 0.693 in the 2-way, 10-shot configuration. In a disease-specific analysis, Hernia obtains the best classification results. Furthermore, Prototypical and Relation Networks demonstrate significantly higher computational efficiency, requiring fewer FLOPs and shorter execution times than MAML and FoMAML. Conclusions: Prototype-based meta-learning, particularly with DenseNet-121, proves to be a robust and computationally efficient approach for few-shot chest X-ray disease classification. These findings highlight its potential for real-world clinical applications, especially in scenarios with limited annotated medical data.