Diagnosing lung cancer typically involves physicians identifying lung nodules in Computed tomography (CT) scans and generating diagnostic reports based on their morphological features and medical expertise. Although advancements have been made in using multimodal large language models for analyzing lung CT scans, challenges remain in accurately describing nodule morphology and incorporating medical expertise. These limitations affect the reliability and effectiveness of these models in clinical settings. Collaborative multi-agent systems offer a promising strategy for achieving a balance between generality and precision in medical applications, yet their potential in pathology has not been thoroughly explored. To bridge these gaps, we introduce LungNoduleAgent, an innovative collaborative multi-agent system specifically designed for analyzing lung CT scans. LungNoduleAgent streamlines the diagnostic process into sequential components, improving precision in describing nodules and grading malignancy through three primary modules. The first module, the Nodule Spotter, coordinates clinical detection models to accurately identify nodules. The second module, the Radiologist, integrates localized image description techniques to produce comprehensive CT reports. Finally, the Doctor Agent System performs malignancy reasoning by using images and CT reports, supported by a pathology knowledge base and a multi-agent system framework. Extensive testing on two private datasets and the public LIDC-IDRI dataset indicates that LungNoduleAgent surpasses mainstream vision-language models, agent systems, and advanced expert models such as GPT-4o, Claude 3.7 Sonnet, LLaMA-3.2 Vision, Qwen2.5-VL, Med-R1, MedGemma, MedAgent-Pro, MedAgents, MDAgent and LLaVA-Med. These results highlight the importance of region-level semantic alignment and multi-agent collaboration in diagnosing nodules. LungNoduleAgent stands out as a promising foundational tool for supporting clinical analyses of lung nodules.

Yanrui Du, Fenglei Fan, Sendong Zhao, Jiawei Cao, Ming Ma, Danyang Zhao, Shuren Qi, Ting Liu, Bing Qin. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

The rapid growth of large models' size has far outpaced that of computing resources. To bridge this gap, encouraged by the parsimonious relationship between genotype and phenotype in the brain's growth and development, we propose the so-called Hyper-Compression that turns the model compression into the issue of parameter representation via a hyperfunction. Specifically, it is known that the trajectory of some low-dimensional dynamic systems can fill the high-dimensional space eventually. Thus, Hyper-Compression, using these dynamic systems as the hyperfunctions, represents the parameters of the target network by their corresponding composition number or trajectory length. This suggests a novel mechanism for model compression, substantially different from the existing pruning, quantization, distillation, and decomposition. Along this direction, we methodologically identify a suitable dynamic system with the irrational winding as the hyperfunction and theoretically derive its associated error bound. Next, guided by our theoretical insights, we propose several engineering twists to make the Hyper-Compression pragmatic and effective. Lastly, systematic and comprehensive experiments on NLP models such as LLaMA and Qwen series and vision models confirm that Hyper-Compression enjoys the following PNAS merits: 1) Preferable compression ratio; 2) No post-hoc retraining; 3) Affordable inference time; and 4) Short compression time. It compresses LLaMA2-7B in an hour and achieves close-to-int4-quantization performance, without retraining and with a performance drop of less than 1%.

Metal artifacts in computed tomography (CT) imaging significantly hinder diagnostic accuracy and clinical decision-making. While deep learning-based metal artifact reduction (MAR) methods have demonstrated promising progress, their clinical application is still constrained by three major challenges: 1) balancing metal artifact reduction with the preservation of critical anatomical structures, 2) effectively capturing the clinical priors of metal artifacts, and 3) dynamically adapting to polychromatic spectral variations. To address these limitations, in this paper, we propose a Self-supervised MAR method for computed Tomography (SMART) that leverages range-null space decomposition (RND) to model metal and tissue LACs separately, and employs implicit neural representation (INR) to learn their respective clinical characteristics without explicit supervision. Specifically, RND decouples metal and tissue LACs into a residual range component for metal LAC modeling, which captures metal artifacts, thus facilitating metal artifact reduction, and a null component for tissue LAC modeling, which focuses on preserving tissue details. To deal with the lack of paired data in clinical settings, we utilize INR to learn the clinical characteristics of these components in a self-supervised manner. Furthermore, SMART incorporates polychromatic spectra into the implicit representation, allowing dynamic adaptation to spectral variations across different imaging conditions. Extensive experiments on one synthetic and two clinical datasets demonstrate the strong potential of SMART in real-world scenarios. By flexibly adapting to spectral variations, it achieves superior generalizability to out-of-distribution clinical data.

Neural networks that overlook the underlying causal relationships among observed variables pose significant risks in high-stakes decision-making contexts due to concerns about the robustness and stability of model performance. To tackle this issue, we present a general approach for embedding hierarchical causal structure among observed variables into a neural network to inform its learning. The proposed methodology, termed causality-informed neural network (CINN), exploits hierarchical causal structure learned from observational data as a structurally informed prior to guide the layer-to-layer architectural design of the neural network while maintaining the orientation of causal relationships in the discovered causal graph. The proposed method involves three steps. First, CINN mines causal relationships from observational data via directed acyclic graph (DAG) learning, where causal discovery is recast as a continuous optimization problem to circumvent the combinatorial nature of DAG learning. Second, we encode the discovered hierarchical causal graph among observed variables into a neural network via a dedicated architecture and loss function. By classifying observed variables in the DAG as root, intermediate, and leaf nodes, we translate the hierarchical causal DAG into CINN by creating a one-to-one correspondence between DAG nodes and certain CINN neurons. For the loss function, both intermediate and leaf nodes in the DAG are treated as target outputs during CINN training, facilitating the co-learning of causal relationships among the observed variables. Finally, as multiple loss components emerge in CINN, we leverage the projection of conflicting gradients (PCGrads) to mitigate the gradient interference among the multiple learning tasks. Computational studies indicate that CINN outperforms several state-of-the-art methods across a broad range of datasets. In addition, an ablation study that incrementally incorporates structural and quantitative causal knowledge into the neural network is conducted to highlight the pivotal role of causal knowledge in enhancing neural network’s prediction performance.

Inspired by neuronal diversity in the biological neural system, a plethora of studies proposed to design novel types of artificial neurons and introduce neuronal diversity into artificial neural networks. Recently proposed quadratic neuron, which replaces the inner-product operation in conventional neurons with a quadratic one, have achieved great success in many essential tasks. Despite the promising results of quadratic neurons, there is still an unresolved issue: Is the superior performance of quadratic networks simply due to the increased parameters or due to the intrinsic expressive capability? Without clarifying this issue, the performance of quadratic networks is always suspicious. Additionally, resolving this issue is reduced to finding killer applications of quadratic networks. In this paper, with theoretical and empirical studies, we show that quadratic networks enjoy parametric efficiency, thereby confirming that the superior performance of quadratic networks is due to the intrinsic expressive capability. This intrinsic expressive ability comes from that quadratic neurons can easily represent nonlinear interaction, while it is hard for conventional neurons. Theoretically, we derive the approximation efficiency of quadratic networks over conventional ones in terms of real space and manifolds. Moreover, from the perspective of the Barron space, we demonstrate that there exists a functional space whose functions can be approximated by quadratic networks in a dimension-free error, but the approximation error of conventional networks is dependent on dimensions. Empirically, experimental results on synthetic data, classic benchmarks, and real-world applications show that quadratic models broadly enjoy parametric efficiency, and the gain of efficiency depends on the task.

Generative artificial intelligence has shown great success in visual content synthesis such that humans struggle to distinguish between real and synthesized images. Forensic research seeks to reveal artifacts in such generated images, ensuring information security or improving generation capability. In this regard, the robustness and interpretability are important for the trustworthy purpose of forensic tasks. However, typical forensic models and their underlying data representations rely on empirical learning algorithms, which cannot effectively handle the high robustness and interpretability requirements beyond experience. As an effective solution, we extend the classical geometric invariants to the forensic research of large-scale generated images. Invariants are handcrafted representations with robust and interpretable geometric principles. However, their discriminability is far from the large scale of today's forensic tasks. We boost the discriminability by extending the classical invariants to the hierarchical architecture of convolutional neural networks. The resulting overcompleteness allows for an automatic selection of task-discriminative features, while retaining the previous advantages of robustness and interpretability. From generative adversarial networks to diffusion models, the forensic with our boosted invariants demonstrates state-of-the-art discriminability against large-scale content diversity. It also exhibits high efficiency on training examples, intrinsic invariance to geometric variations, and better interpretability of the forensic process.

Motorimagery EEG classification plays a crucial role in non-invasive Brain-Computer Interface (BCI) research. However, the performance of classification is affected by the non-stationarity and individual variations of EEG signals. Simply pooling EEG data with different statistical distributions to train a classification model can severely degrade the generalization performance. To address this issue, the existing methods primarily focus on domain adaptation, which requires access to the test data during training. This is unrealistic and impractical in many EEG application scenarios. In this paper, we propose a novel multi-source domain generalization framework called EEG-DG, which leverages multiple source domains with different statistical distributions to build generalizable models on unseen target EEG data. We optimize both the marginal and conditional distributions to ensure the stability of the joint distribution across source domains and extend it to a multi-source domain generalization framework to achieve domain-invariant feature representation, thereby alleviating calibration efforts. Systematic experiments conducted on a simulative dataset, BCI competition IV 2a, 2b, and OpenBMI datasets, demonstrate the superiority and competitive performance of our proposed framework over other state-of-the-art methods. Specifically, EEG-DG achieves average classification accuracies of 81.79% and 87.12% on datasets IV-2a and IV-2b, respectively, and 78.37% and 76.94% for inter-session and inter-subject evaluations on dataset OpenBMI, which even outperforms some domain adaptation methods.

Computed tomography (CT) stands as a ubiquitous medical diagnostic tool. Nonetheless, the radiation-related concerns associated with CT scans have raised public apprehensions. Mitigating radiation dosage in CT imaging poses an inherent challenge as it inevitably compromises the fidelity of CT reconstructions, impacting diagnostic accuracy. While previous deep learning techniques have exhibited promise in enhancing CT reconstruction quality, they remain hindered by the reliance on paired data, which is arduous to procure. In this study, we present a novel approach named Variational Score Solver (VSS) for sparse-view reconstruction without paired data. Our approach entails the acquisition of a probability distribution from densely sampled CT reconstructions, employing a latent diffusion model. High-quality reconstruction outcomes are achieved through an iterative process, wherein the diffusion model serves as the prior term, subsequently integrated with the data consistency term. Notably, rather than directly employing the prior diffusion model, we distill prior knowledge by finding the fixed point of the diffusion model. This framework empowers us to exercise precise control over the process. Moreover, we depart from modeling the reconstruction outcomes as deterministic values, opting instead for a distribution-based approach. This enables us to achieve more accurate reconstructions utilizing a trainable model. Our approach introduces a fresh perspective to the realm of zero-shot CT reconstruction, circumventing the constraints of supervised learning. Extensive qualitative and quantitative experiments unequivocally demonstrate that VSS surpasses other contemporary unsupervised and achieves comparable results compared to the most advanced supervised methods in sparse-view reconstruction tasks. Codes are available in https://github.com/fpsandnoob/vss.

Learning a generalizable medical image segmentation model is an important but challenging task since the unseen (testing) domains may have significant discrepancies from seen (training) domains due to different vendors and scanning protocols. Existing segmentation methods, typically built upon domain generalization (DG), aim to learn multi-source domain-invariant features through data or feature augmentation techniques, but the resulting models either fail to characterize global domains during training or cannot sense unseen domain information during testing. To tackle these challenges, we propose a domain Unifying and Adapting network (UniAda) for generalizable medical image segmentation, a novel "unifying while training, adapting while testing" paradigm that can learn a domain-aware base model during training and dynamically adapt it to unseen target domains during testing. First, we propose to unify the multi-source domains into a global inter-source domain via a novel feature statistics update mechanism, which can sample new features for the unseen domains, facilitating the training of a domain base model. Second, we leverage the uncertainty map to guide the adaptation of the trained model for each testing sample, considering the specific target domain may be outside the global inter-source domain. Extensive experimental results on two public cross-domain medical datasets and one in-house cross-domain dataset demonstrate the strong generalization capacity of the proposed UniAda over state-of-the-art DG methods. The source code of our UniAda is available at https://github.com/ZhouZhang233/UniAda.

Inspired by the diversity of biological neurons, quadratic artificial neurons can play an important role in deep learning models. The type of quadratic neurons of our interest replaces the inner-product operation in the conventional neuron with a quadratic function. Despite promising results so far achieved by networks of quadratic neurons, there are important issues not well addressed. Theoretically, the superior expressivity of a quadratic network over either a conventional network or a conventional network via quadratic activation is not fully elucidated, which makes the use of quadratic networks not well grounded. In practice, although a quadratic network can be trained via generic backpropagation, it can be subject to a higher risk of collapse than the conventional counterpart. To address these issues, we first apply the spline theory and a measure from algebraic geometry to give two theorems that demonstrate better model expressivity of a quadratic network than the conventional counterpart with or without quadratic activation. Then, we propose an effective training strategy referred to as referenced linear initialization (ReLinear) to stabilize the training process of a quadratic network, thereby unleashing the full potential in its associated machine learning tasks. Comprehensive experiments on popular datasets are performed to support our findings and confirm the performance of quadratic deep learning. We have shared our code in https://github.com/FengleiFan/ReLinear.

A neural network (NN) with the widely-used ReLU activation has been shown to partition the sample space into many convex polytopes for prediction. However, the parametric way a NN and other machine learning models use to partition the space has imperfections, e.g., the compromised interpretability for complex models, the inflexibility in decision boundary construction due to the generic character of the model, and the risk of being trapped into shortcut solutions. In contrast, although the nonparameterized models can adorably avoid or downplay these issues, they are usually insufficiently powerful either due to over-simplification or the failure to accommodate the manifold structures of data. In this context, we first propose a new type of machine learning models referred to as Manifoldron that directly derives decision boundaries from data and partitions the space via manifold structure discovery. Then, we systematically analyze the key characteristics of the Manifoldron such as manifold characterization capability and its link to NNs. The experimental results on four synthetic examples, 20 public benchmark datasets, and one real-world application demonstrate that the proposed Manifoldron performs competitively compared to the mainstream machine learning models. We have shared our code in https://github.com/wdayang/Manifoldron for free download and evaluation.

The accurate detection of sperms and impurities is a very challenging task, facing problems such as the small size of targets, indefinite target morphologies, low contrast and resolution of the video, and similarity of sperms and impurities. So far, the detection of sperms and impurities still largely relies on the traditional image processing and detection techniques which only yield limited performance and often require manual intervention in the detection process, thus unfavorably escalating the time cost and injecting the subjective bias into the analysis. Encouraged by the success of deep learning methods in numerous object detection tasks, here we report a deep learning network: ACTIVE based on Double Branch Feature Extraction Network (DBFEN) and Cross-conjugate Feature Pyramid Network (CCFPN). DBFEN extracts visual features from tiny objects with a double branch structure, and CCFPN fuses the features extracted by DBFEN to enhance the description of the position and high-level semantic information. Our work is the pioneer of introducing deep learning approaches to the detection of sperms and impurities. Experiments show that the highest AP50of the sperm and impurity detection is 91.13% and 59.64%, which lead its competitors by a substantial margin and establish the state-of-the-art results in this problem. Code is available for readers’ free evaluation at https://github.com/anheqiao-neu/ACTIVE.

Many contemporary studies utilize grid-based models for neural field representation, but a systematic analysis of grid-based models is still missing, hindering the improvement of those models. Therefore, this paper introduces a theoretical framework for grid-based models. This frame-work points out that these models' approximation and generalization behaviors are determined by grid tangent ker-nels (GTK), which are intrinsic properties of grid-based models. The proposed framework facilitates a consistent and systematic analysis of diverse grid-based models. Furthermore, the introduced framework motivates the development of a novel grid-based model named the Multiplicative Fourier Adaptive Grid (MulFAGrid). The numerical analysis demonstrates that MulFAGrid exhibits a lower generalization bound than its predecessors, indicating its robust generalization performance. Empirical studies reveal that MulFAGrid achieves state-of-the-art performance in various tasks, including 2D image fitting, 3D signed distance field (SDF) reconstruction, and novel view synthesis, demonstrating superior representation ability. The project website is available at this link.

The vulnerability of deep neural networks to adversarial perturbations has been widely perceived in the computer vision community. From a security perspective, it poses a critical risk for modern vision systems, e.g., the popular Deep Learning as a Service (DLaaS) frameworks. For protecting deep models while not modifying them, current algorithms typically detect adversarial patterns through discriminative decomposition for natural and adversarial data. However, these decompositions are either biased towards frequency resolution or spatial resolution, thus failing to capture adversarial patterns comprehensively. Also, when the detector relies on few fixed features, it is practical for an adversary to fool the model while evading the detector (i.e., defense-aware attack). Motivated by such facts, we propose a discriminative detector relying on a spatial-frequency Krawtchouk decomposition. It expands the above works from two aspects: 1) the introduced Krawtchouk basis provides better spatial-frequency discriminability, capturing the differences between natural and adversarial data comprehensively in both spatial and frequency distributions, w.r.t. the common trigonometric or wavelet basis; 2) the extensive features formed by the Krawtchouk decomposition allows for adaptive feature selection and secrecy mechanism, significantly increasing the difficulty of the defense-aware attack, w.r.t. the detector with few fixed features. Theoretical and numerical analyses demonstrate the uniqueness and usefulness of our detector, exhibiting competitive scores on several deep models and image sets against a variety of adversarial attacks.

Filling gaps in high-resolution satellite imagery is essential for tracking vegetation changes over time. Spatiotemporal fusion (STF) aims to create fusion products that improve both spatial resolution and temporal coverage by using images from various remote sensing sources. However, most existing STF methods rely on the assumption that reflectance values for the same land-cover type remain constant between base and prediction dates, a premise often invalidated by the variability in vegetation disturbance and recovery scenarios, where differences in disturbance intensity, patterns, and phenological stages challenge this uniformity. Therefore, we propose a novel Transformer-based method, the spatiotemporal integration network (STINet), which is effective in fusing multiscale spatiotemporal dynamic features. STINet is structured around three key components. The feature fusion (FF) block effectively integrates multiscale spatiotemporal information into a deep learning (DL) framework. The adaptive feature extraction (AFE) block significantly improves the precision of pixel-level features, essential for detecting subtle changes in diverse vegetation patterns. The spatiotemporal-wise multihead self-attention (ST-MSA) module through its innovative self-attention mechanism across spatiotemporal dimensions, facilitated the reconstruction of vegetation dynamics. To verify the effectiveness and robustness of the proposed method, we conducted experiments in three carefully selected scenarios using multisensor and multitemporal imagery to reconstruct the dynamic changes in vegetation due to various disturbances and recovery processes. Compared to the four typical fusion methods [enhanced spatial and temporal adaptive reflectance fusion model (ESTARFM), flexible spatiotemporal data fusion method (FSDAF), extended super-resolution convolutional neural network (ESRCNN), and multiscene spatiotemporal fusion network (MUSTFN)], STINet achieved the best performance in preserving both spatial texture and spectral value. Furthermore, we showcased the applicability and effectiveness of STINet in accurately capturing phenological changes and distinguishing various farming activities.

The Retinex model is one of the most representative and effective methods for low-light image enhancement. However, the Retinex model does not explicitly tackle the noise problem and shows unsatisfactory enhancing results. In recent years, due to the excellent performance, deep learning models have been widely used in low-light image enhancement. However, these methods have two limitations. First, the desirable performance can only be achieved by deep learning when a large number of labeled data are available. However, it is not easy to curate massive low-/normal-light paired data. Second, deep learning is notoriously a black-box model. It is difficult to explain their inner working mechanism and understand their behaviors. In this article, using a sequential Retinex decomposition strategy, we design a plug-and-play framework based on the Retinex theory for simultaneous image enhancement and noise removal. Meanwhile, we develop a convolutional neural network-based (CNN-based) denoiser into our proposed plug-and-play framework to generate a reflectance component. The final image is enhanced by integrating the illumination and reflectance with gamma correction. The proposed plug-and-play framework can facilitate both post hoc and ad hoc interpretability. Extensive experiments on different datasets demonstrate that our framework outcompetes the state-of-the-art methods in both image enhancement and denoising.

Recent years have witnessed an increasing interest in the correspondence between infinitely wide networks and Gaussian processes. Despite the effectiveness and elegance of the current neural network Gaussian process theory, to the best of our knowledge, all the neural network Gaussian processes (NNGPs) are essentially induced by increasing width. However, in the era of deep learning, what concerns us more regarding a neural network is its depth as well as how depth impacts the behaviors of a network. Inspired by a width-depth symmetry consideration, we use a shortcut network to show that increasing the depth of a neural network can also give rise to a Gaussian process, which is a valuable addition to the existing theory and contributes to revealing the true picture of deep learning. Beyond the proposed Gaussian process by depth, we theoretically characterize its uniform tightness property and the smallest eigenvalue of the Gaussian process kernel. These characterizations can not only enhance our understanding of the proposed depth-induced Gaussian process but also pave the way for future applications. Lastly, we examine the performance of the proposed Gaussian process by regression experiments on two benchmark datasets.

While classic studies proved that wide networks allow universal approximation, recent research and successes of deep learning demonstrate the power of deep networks. Based on a symmetric consideration, we investigate if the design of artificial neural networks should have a directional preference, and what the mechanism of interaction is between the width and depth of a network. Inspired by the De Morgan law, we address this fundamental question by establishing a quasi-equivalence between the width and depth of ReLU networks. We formulate two transforms for mapping an arbitrary ReLU network to a wide ReLU network and a deep ReLU network respectively, so that the essentially same capability of the original network can be implemented. Based on our findings, a deep network has a wide equivalent, and vice versa, subject to an arbitrarily small error.

Image denoising is a prerequisite for downstream tasks in many fields. Low-dose and photon-counting computed tomography (CT) denoising can optimize diagnostic performance at minimized radiation dose. Supervised deep denoising methods are popular but require paired clean or noisy samples that are often unavailable in practice. Limited by the independent noise assumption, current self-supervised denoising methods cannot process correlated noises as in CT images. Here we propose the first-of-its-kind similarity-based self-supervised deep denoising approach, referred to as Noise2Sim, that works in a nonlocal and nonlinear fashion to suppress not only independent but also correlated noises. Theoretically, Noise2Sim is asymptotically equivalent to supervised learning methods under mild conditions. Experimentally, Nosie2Sim recovers intrinsic features from noisy low-dose CT and photon-counting CT images as effectively as or even better than supervised learning methods on practical datasets visually, quantitatively and statistically. Noise2Sim is a general self-supervised denoising approach and has great potential in diverse applications.

Inspired by complexity and diversity of biological neurons, our group proposed quadratic neurons by replacing the inner product in current artificial neurons with a quadratic operation on input data, thereby enhancing the capability of an individual neuron. Along this direction, we are motivated to evaluate the power of quadratic neurons in popular network architectures, simulating human-like learning in the form of "quadratic-neuron-based deep learning". Our prior theoretical studies have shown important merits of quadratic neurons and networks in representation, efficiency, and interpretability. In this paper, we use quadratic neurons to construct an encoder-decoder structure, referred as the quadratic autoencoder, and apply it to low-dose CT denoising. The experimental results on the Mayo low-dose CT dataset demonstrate the utility and robustness of quadratic autoencoder in terms of image denoising and model efficiency. To our best knowledge, this is the first time that the deep learning approach is implemented with a new type of neurons and demonstrates a significant potential in the medical imaging field.

With the increasing of economy, China had experienced a very fast land use change, especially farmland loss. Lots of farmland loss resulted in degeneration of ecosystem and affected the economic development deeply, so there is a great need to monitor the farmland loss and analysis their economic implication. One of most common methods to monitor farmland loss is remote sensing data because of its fast, dynamic characteristics. This paper analysis the relation between farmland loss and economy in Guangzhou Municipality from 1998 to 2003, and an evaluation model is found in order to discuss more detail problems and the implications to economy. Results show that the farmland loss (from 1998 to 2003) of Guangzhou Municipality is very serious and farmland economic transform efficiency is low generally.

Shenzhen is one of the cities with the most rapid urbanization speed in past 20 years in the world. The study of urbanization in Shenzhen can provide valuable information for sustainable development to the other cities in Pearl River Delta areas, in China and even in the world. Although many previous studies demonstrated the land-use changes by remote sensing, the results tire not systematic due to the different boundaries and time durations. For researching the urban sprawl, this paper extracted the urban-used land by Landsat TM/ETM data, detected the urban land change from 1998 to 2003 and analyzed their spatial distributions and the economic implications.