Unpaired learning is a vital paradigm for singleimage laparoscopic desmoking, addressing the common challenge of clinical data scarcity. However, prevailing methods within this paradigm often fail by producing trivial solutions, where the output is nearly identical to the smoky input. We identify that this failure stems from an anomalous penalty mechanism created by the conflict between cycle-consistency loss and inherent label noise in the training data. To address this fundamental limitation, we propose a novel Decoupled Pseudo-Label Refinement (DPLR) framework. This framework utilizes a teacher-student architecture wherein a teacher model first generates pseudopaired data. This data then undergoes a systematic refinement pipeline to filter out noisy and trivial samples before being used to train a student model via supervised learning. Furthermore, an adaptive inference strategy is proposed to leverage the complementary strengths of both models during deployment. Extensive experiments demonstrate our method surpasses state-of-the-art approaches in both quantitative metrics and visual quality on the public DeSmoke-LAP dataset. The model also demonstrates strong generalization in zero-shot evaluations on multiple unseen datasets, suggesting it learns generalizable desmoking knowledge rather than dataset-specific features. The framework's success as an enhancement paradigm for various baseline models further underscores its potential for broader applications in unpaired medical image restoration. Code at https://anonymous.4open.science/r/DPLR-6C8C is available.

Automated radiology report generation is essential in clinical practice. However, diagnosing radiological images typically requires physicians 5-10 minutes, resulting in a waste of valuable healthcare resources. Existing studies have not fully leveraged knowledge from historical radiology reports, lacking sufficient and accurate prior information. To address this, we propose a Topic-Keyword Semantic Guidance (TKSG) framework. This framework uses BiomedCLIP to accurately retrieve historical similar cases. Supported by multimodal, TKSG accurately detects topic words (disease classifications) and keywords (common symptoms) in diagnoses. The probabilities of topic terms are aggregated into a topic vector, serving as global information to guide the entire decoding process. Additionally, a semantic-guided attention module is designed to refine local decoding with keyword content, ensuring report accuracy and relevance. Experimental results show that our model achieves excellent performance on both IU X-Ray and MIMIC-CXR datasets. The code is available at https://github.com/SCNU203/TKSG

Knowledge Tracing (KT) serves as a fundamental component of Intelligent Tutoring Systems (ITS), enabling these systems to monitor and understand learners’ progress by modeling their knowledge state. However, many existing KT models primarily focus on fitting the sequences of learners’ interactions, and often overlook the knowledge state itself. This limitation leads to reduced interpretability and insufficient instructional support from the ITS. To address this challenge, we propose AlignKT, which employs a frontend-to-backend architecture to explicitly model a stable knowledge state. In this approach, the preliminary knowledge state is aligned with an additional criterion. Specifically, we define an ideal knowledge state based on pedagogical theories as the alignment criterion, providing a foundation for interpretability. We utilize five encoders to implement this set-up, and incorporate a contrastive learning module to enhance the robustness of the alignment process. Through extensive experiments, AlignKT demonstrates superior performance, outperforming seven KT baselines on three real-world datasets. It achieves state-of-the-art results on two of these datasets and exhibits competitive performance on the third. The code of this work is available at https://github.com/SCNU203/AlignKT.

Accurate segmentation of liver tumors is crucial for their proper diagnosis and treatment. However, achieving high levels of precision typically depends on meticulous manual annotation, a process that is not only labor-intensive but also constrained by the scarcity of large-scale, real-world datasets. These datasets are indispensable for the training and validation of segmentation algorithms. Furthermore, the liver’s considerable variation in size, shape, and pathology type poses a challenge in collecting a sufficient number of image samples that adequately represent this diversity. To surmount the challenges of the time-consuming manual annotation process and the limitations in data acquisition, there is an urgent need for the development of an efficient and comprehensive tumor generation method. Some existing approaches, such as those utilizing Gaussian blurred ellipses to simulate tumors, fail to accurately reflect the biological complexity and pathological diversity of liver tumors. In this article, we introduce an innovative tumor synthesis method Latent Diffusion Model for Pathology (LDMP) that leverages medical imaging knowledge to more accurately replicate the intricacies of liver tumor morphology and pathology. This approach aims to enhance the quality and diversity of training data, thereby improving the performance of segmentation algorithms and ultimately contributing to more precise diagnoses and treatments. The method uses deep learning techniques, particularly diffusion models, to simulate real liver CT images and incorporates the biological properties of tumors into the synthesis process to generate realistic tumor images. The quality of the synthetic images is assessed using Principal Component Analysis (PCA) and Kullback–Leibler (KL) divergence to ensure the authenticity of the tumor’s spatial structure. Experimental results show that the proposed method can significantly improve the Dice Similarity Coefficient (DSC) of the tumor segmentation model and enable researchers to freely define the size and blur degree of the tumor, thereby creating medical images with precise annotations. In addition, we introduce a self-checking step before the output of synthetic data, which provides a new paradigm in the field of image synthesis and effectively compensates for potential errors in synthetic data. Our approach not only provides an effective solution for medical image analysis but also provides high-quality synthetic image resources for medical education and clinical practice. Codes are available at: https://github.com/jhf0721/tumor .

Super-resolution combined with novel image synthesis is an advanced image processing method to synthesize low-resolution images into new high-resolution images. NeRF-SR is the first model to obtain decent multi-view super-resolution results with only low-resolution input images, but the super-sampling method implemented using the original Nerf’s MLP network cannot represent the complex details of the scene well. We consider that the volume density and color features obtained by the MLP network do not take into account the global geometry along the ray and the color relationship between the sampling points. To tackle this challenge, we introduce an attention-based model and auto-encoding network to synthesize high-fidelity views from low-resolution input to high-resolution output. The attention-based model mixes the pixel color information of the sampling points on each ray and supervises using ground-truth colors. At the same time, the auto-encoding network learns the global geometry along the ray. Experimental results demonstrate that our model can produce high-quality results for high-resolution new view synthesis, both on synthetic and real-world datasets.

Sleep stage classification is significant for sleep specialists to evaluate sleep quality and diagnose sleep disorders. Machine learning and deep learning technologies are widely employed to build automatic sleep stage classification models. However, how to make full use of and integrate multiple heterogeneous information such as unlabeled information, topological information, frequency information and neighboring information to improve classification is still an open problem. To address this issue, we propose a multi-stream fusion network named MSF-SleepNet for sleep stage classification, with contrastive learning to combine spatial, temporal, and spectral features. Firstly, we design a contrastive learning framework to learn general features from unlabeled information. On this basis, we apply graph structure learning, Chebyshev graph convolution and temporal convolution to capture spatial-temporal features from topological information of human body in non-Euclidean space and transition rules among sleep stages. Secondly, we utilize the short-time Fourier transform and Gate Recurrent Unit to gain spectral-temporal features from frequency information of different time series signals in Euclidean space and neighboring information of adjacent signal segments. Finally, fusing spatial-temporal features and spectral-temporal features can further enhance the performance of sleep stage classification. Experimental results on publicly available datasets of ISRUC-S3 show that our method is more effective in integrating heterogeneous information and achieves better performance than existing state-of-the-art methods.

Cross-modal retrieval has gained much attention due to the growing demand for enormous multi-modal data in recent years (i.e., image-text or text-image retrieval). In order to alleviate the problem of ignoring the existence of irrelevant information between images and texts, this paper proposes Deep Pairwise Ranking model with multi-label information for Cross-Modal retrieval (DPRCM). DPRCM directly learns a mapping from images and texts to a compact Euclidean space where distances correspond to the similarity measure of images and texts. The bi-triplet loss function in DPRCM reduces the distance between associated images and texts on the common subspace and increases the margin of independent samples. The classification loss function can better utilize the multi-label information to reduce the semantic gap between image features and text descriptions. Experiments on three widely-used datasets show that DPRCM can achieve competitive performance compared to state-of-the-art methods.

Scene text detection is one of the most challenging tasks in many computer vision applications due to the large variety of scene text appearance and the complexity of scene context. In this paper, we propose an end-to-end trainable framework RefineText for multi-oriented scene text detection, which has a strong ability to detect different-scale texts and split them precisely. High-resolution semantic features are first generated by our designed Feature Refinement Module, which refines features progressively at multiple levels of abstraction. Then text regions are densely produced on the high-level semantic features and followed by Non-Maximum Suppression(NMS) to get final detection results. Experiments on benchmark datasets including ICDAR 2015, ICDAR 2013 and MSRA TD500 demonstrate that our proposed method has competitive performance and strong robustness.

Multi-modal data lies on heterogeneous feature spaces, which brings a significant challenge to cross-modal retrieval. Some works have been proposed to cope with this problem by learning a common subspace. However, previous methods often learn the common subspace by enhancing the relation between embedded features and relevant class labels but ignore the relation between embedded features and irrelevant class labels. Additionally, most methods assume that irrelevant samples are of equal importance. Considering this, we propose to train an optimal common embedding space via cross-modal learning to rank with adaptive listwise constraint (CMAL2R) based on two-branch neural networks. The listwise loss function in CMAL2R adaptively assigns larger margins to harder irrelevant samples, strengthening the relation between embedded features and irrelevant class labels. Experiments on Wikipedia and Pascal datasets demonstrate the effectiveness for bi-directional image-text retrieval.

The Software Project Scheduling Problem (SPSP) focuses on the management of software engineers and tasks in a software project so as to complete the tasks with a minimal cost and duration. It's becoming more and more important and challenging with the rapid development of software industry. In this paper, we employ a Multi-objective Evolutionary Algorithm using Decomposition and Ant Colony (MOEA/D-ACO) to solve the SPSP. To the best of our knowledge, it is the first application of Multi-objective Ant Colony Optimization (MOACO) to SPSP. Two heuristics capable of guiding the algorithm to search better in the SPSP model are examined. Experiments are conducted on a set of 36 publicly available instances. The results are compared with the implementation of another multi-objective evolutionary algorithm called NSGA-II for SPSP. MOEA/D-ACO does not outperform NSGA-II for most of complex instances in terms of Pareto Front. But MOEA/D-ACO can obtain solutions with much less time for all instances in our experiments and it outperforms NSGA-II with less duration for most of test instances. The performance may be improved with tuning of the algorithm such as incorporating more heuristic information or using other MOACO algorithms, which deserve further investigation.

As one of the most challenging combinatorial optimization problems in scheduling, the resource-constrained project sche-duling problem (RCPSP) has attracted numerous scholars' interest resulting in considerable research in the past few decades. However, most of these papers focused on the single objective RCPSP; only a few papers concentrated on the multi-objective resource-constrained project scheduling problem (MORCPSP). Inspired by a procedure called electromagnetism (EM), which can help a generic population-based evolutionary search algorithm to obtain good results for single objective RCPSP, in this paper we attempt to extend EM and hybridize it with three reputable state-of-the-art multi-objective evolutionary algorithms (MOEAs) i.e. NSGA-II, SPEA2 and MOEA/D, for MORCPSP. Our two objectives are minimizing makespan and total tardiness. We perform computational experiments on standard benchmark datasets. Empirical comparison and analysis of the results obtained by the hybridization versions of EM with NSGA-II, SPEA2 and MOEA/D are conducted. The results demonstrate that EM can improve the performance of NSGA-II and SPEA2.

Computer-assisted testing systems are promising in generating tests efficiently and effectively for evaluating a person's skill. This paper develops a novel intelligent testing system for both teachers and students. Based on the browser/server structure, the proposed testing system comprises a question bank and five modules, offering the features of self-adaptation, reliability, and flexibility for generating parallel tests with identical test ability. The core of the developed system is the ant-colony-optimization-based test composition (ACO-TC) method, which aims at generating high-quality tests for examinations and satisfying multiple requirements. As an advanced computational intelligence algorithm, the proposed ACO-TC method uses a colony of ants to select appropriate questions from a question bank to construct solutions. Pheromone and heuristic information is designed for facilitating the ants' selection. The system is analyzed by composing tests in different situations. The generated tests not only match the expected total completion time, the concept proportions, the average difficulty, and the score proportions of different question types, but also have high average discrimination degrees of questions. The experimental results also show that the system can always generate high-quality tests from question banks with various sizes.

The conventional k-means clustering algorithm must know the number of clusters in advance and the clustering result is sensitive to the selection of the initial cluster centroids. The sensitivity may make the algorithm converge to the local optima. This paper proposes an improved k-means clustering algorithm based on quantum-inspired genetic algorithm (KMQGA). In KMQGA, Q-bit based representation is employed for exploration and exploitation in discrete 0-1 hyperspace by using rotation operation of quantum gate as well as three genetic algorithm operations (selection, crossover and mutation) of Q-bit. Without knowing the exact number of clusters beforehand, the KMQGA can get the optimal number of clusters as well as providing the optimal cluster centroids after several iterations of the four operations (selection, crossover, mutation, and rotation). The simulated datasets and the real datasets are used to validate KMQGA and to compare KMQGA with an improved k-means clustering algorithm based on the famous variable string length genetic algorithm (KMVGA) respectively. The experimental results show that KMQGA is promising and the effectiveness and the search quality of KMQGA is better than those of KMVGA.

Research into setting the values of the acceleration coefficients c1and c2in Particle Swarm Optimization (PSO) is one of the most significant and promising areas in evolutionary computation. Parameters c1and c2in PSO indicate the “self-cognitive” and “social-influence” components which are important for the ability to explore and converge respectively. Instead of using fixed value of c1and c2with 2.0, this paper presents the use of clustering analysis to adaptively adjust the value of these two parameters in PSO. By applying the K-means algorithm, distribution of the population in the search space is clustered in each generation. An adaptive system which is based on considering the relative size of the cluster containing the best particle and the one containing the worst particle is used to adjust the values of c1and c2. The proposed method has been applied to optimize multidimensional mathematical functions, and the simulation results demonstrate that the proposed method performs with a faster convergence rate and better solutions when compared with the methods with fixed values of c1and c2.