He Geng, Yangmin Huang, Lixian Lai, Qianyun Du, Hui Chu, Zhiyang He, Jiaxue Hu, Xiaodong Tao. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

The integration of multi-omics data, including DNA methylation, mRNA expression, and miRNA profiles, is crucial for accurate disease subtyping and outcome prediction in complex disorders such as Alzheimer's disease and various cancers. However, the inherent heterogeneity and inconsistency among omics views present significant challenges for reliable data fusion. To address these issues, we propose KANTrust, a novel framework for trustworthy multi-omics classification that explicitly models both epistemic and aleatoric uncertainties. Our method combines a Kolmogorov-Arnold Network (KAN)enhanced robust representation module, a contrastive evidence consistency module, and an evidence-theoretic fusion module to achieve reliable multi-view integration. KANTrust adaptively highlights informative features within each omics modality, promotes semantic alignment across views, and quantifies uncertainty through a Dempster-Shafer framework. Experimental evaluations on four real-world biomedical datasets demonstrate that KANTrust consistently outperforms state-of-the-art methods in both binary and multi-class classification tasks. Code is available at https://github.com/wcj6/KANTrust.

The concept bottleneck model (CBM), as a technique improving interpretability via linking predictions to human-understandable concepts, makes high-risk and life-critical medical image classification credible. Typically, existing CBM methods associate the final layer of visual encoders with concepts to explain the model’s predictions. However, we empirically discover the phenomenon of concept preference variation, that is, the concepts are preferably associated with the features at different layers than those only at the final layer; yet a blind last-layer-based association neglects such a preference variation and thus weakens the accurate correspondences between features and concepts, impairing model interpretability. To address this issue, we propose a novel Multi-layer Visual Preference-enhanced Concept Bottleneck Model (MVP-CBM), which comprises two key novel modules: (1) intra-layer concept preference modeling, which captures the preferred association of different concepts with features at various visual layers, and (2) multi-layer concept sparse activation fusion, which sparsely aggregates concept activations from multiple layers to enhance performance. Thus, by explicitly modeling concept preferences, MVP-CBM can comprehensively leverage multi-layer visual information to provide a more nuanced and accurate explanation of model decisions. Extensive experiments on several public medical classification benchmarks demonstrate that MVP-CBM achieves state-of-the-art accuracy and interoperability, verifying its superiority. Code is available at https://github.com/wcj6/MVP-CBM.

The advancement of Zero-Shot Learning in the medi-cal domain has been driven forward by using pretrained models on large-scale image-text pairs, focusing on image-text alignment. However, existing methods primarily rely on cosine similarity for alignment, which may not fully capture the complex relationship between medical images and reports. To address this gap, we introduce a novel approach called Cross-Attention Alignment for Radiology Zero-Shot Classification (CARZero). Our approach innovatively leverages cross-attention mechanisms to process image and report features, creating a Similarity Representation that more accurately reflects the intricate relationships in medical semantics. This representation is then linearly projected to form an image-text similarity matrix for cross-modality alignment. Additionally, recognizing the pivotal role of prompt selection in zero-shot learning, CARZero in-corporates a Large Language Model-based prompt alignment strategy. This strategy standardizes diverse diagnostic expressions into a unified format for both training and inference phases, overcoming the challenges of manual prompt design. Our approach is simple yet effective, demonstrating state-of-the-art performance in zero-shot classification on five official chest radiograph diagnostic test sets, including remarkable results on datasets with long-tail distributions of rare diseases. This achievement is attributed to our new image-text alignment strategy, which effectively addresses the complex relationship between medical images and reports. Code and models are available at https://github.com/laihaoran/CARZero.

Considering a poor condition in using traditional instruments to monitor sea state near Beibu Gulf, this paper presents an economic, flexible, low cost and high accuracy GPS buoy which is mainly used to verify wave parameters which are produced from SAR image inversion. Tide information is removed by wavelet transform, and wind wave information is used to calculate spectrum by Fast Fourier Transform (FFT). In the experiment, GPS buoy presents a flexible and low cost performance in monitoring wind wave. The results show that calculated wave parameters can reach a satisfied and acceptable accuracy and can be used in verifying the outcomes which are calculated by the SAR.

Most of the surface soil moisture (SSM) models developed in recent decades are not for karst area where the surface soil is thin. Consequently, this paper presents a novel algorithm for retrieval of the SSM on the basis of “Optical Vegetation Coverage” for CBERS-02B imagery. Jili village in a typical karst area is chosen as the study area, and the retrieved SSM by Landsat TM satellite imagery is chosen for evaluating the accuracy of the proposed model. It shows that the relative accuracy of the mean SSM overall reachs up to 91.26%, and the correlation coefficient R is up to 0.8. Moreover, the R of the rocky desertification land and the dry land even reaches more than 0.9. With these experimental results, it can be demonstrated that the proposed model based on CBERS-02B satellite imagery has a high precision in the retrieval of SSM in karst area.

Inversing the surface wave spectrum from SAR (Synthetic Aperture Radar) is widely used to obtain the ocean wave parameters in a large ocean area. Considering that the ocean wave is affected by the water depth factor in coastalarea, this study focuses on the inversion of wave parameters in shallow ocean area from SAR images, and using the shallow water TMA spectrum as the first-guess spectrum to develop the inversion method presented by Hasselmann. The surface wave spectrum under swell-dominated case can be inversed from SAR image spectrum directly.So this paper selects 10 sub-images of one ERS-2 image which locates in Zhanjiang city sea area where under the wind-wave dominated cases. The wave lengths and wave directions are calculated from the inversion wave spectrums.Radon transform method is used to compare results. The result indicates that the wave lengths calculated by this inversion method and the radon transform agree with the bias of 4.07m, the standard deviation of 5.05m and the correlation coefficient of 0.953; the wave directions calculated by this inversion method and the radon transform agree with the bias of 1.93°; the standard deviation of 2.70°and the correlation coefficient of 0.955. When the water depth decreases, the change of wave length calculated by this inversion method is consistent with the change of ocean wave length.

The method for retrieving the ocean wave spectrum from the Synthetic Aperture Radar (SAR) imagery is currently the most widely used for calculation of the wave high, wavelength and wave direction. However the most current of methods are relatively complicated. This paper presents an algorithm which is based on Radon transform. In this algorithm, firstly, with the property of Radon transform that the line in the SAR image and the point in the transformed space are one-to-one correspondence, the texture feature of the wave in SAR image is detected to calculate the wavelength and probable wave direction. Secondly, according to the theory of the ocean wave and the trends of the wavelength in two close sub-images, the actual wave direct ion is determined eventually. This article selects an ERS - 2 SAR image around Taiwan as the study area. The results of calculation are compared with the results obtained by Two-Dimensional Fast Fourier Transform (2D- FFT). The study results demonstrate that the use of the Radon transform is available without loss of the accuracy when retrieving the wave wavelength and wave direction from the ERS-2 SAR images.

This paper presents the study of SAR imaging spectrum simulation using the shallow water TMA (Texel-Marsen-Arsloe) spectrum. The study area is located in Beibu Gulf of the South China Sea. The simulated study first selects coastal sea wave spectra with different water depths ranging from 10 m through 40 m to analyze the influence of water depth on SAR imaging spectrum. Furthermore, the change of SAR imaging spectrum is compared by changing the direction of waves propagation at 0°, 45° and 90°, wind speed at 10 m/s through 18 m/s and polarization modes with HH and VV at a certain water depth with an average water depth of 30m. The results discover when the water depth decreases, the SAR spectra peak moves to the high wavenumber region.

This paper presents a new generation of scannerless laser radar measurement system, called GLidar in our project. This system does not require the scanning device; as a result, the dimension of the entire measuring system is small, lightweight, and reliability. The proposed scannerless LiDAR system is especially designated for a small civil UAV platform under a low altitude operation. This paper presents the principle of the array LiDAR imaging, mathematical models of calculating the 3-D coordinates of the laser radar footprints with respect to a mapping coordinate system. Some simulated results are presented on the basis of a test field located in Virginia Wytheville, USA. The simulated results demonstrated that the designated new generation of array LiDAR can achieve 0.06–0.10 m in flat area, 0.31–0.62 m in the edge of house, and 1.23–1.78 m in the forested area compared to an existing DSM data.

This paper describes a new, physically interpretable, fully automatic algorithm for removal of tissue autofluorescence (AF) from fluorescence microscopy images, by non-negative matrix factorization. Measurement of signal intensities from the concentration of certain fluorescent reporter molecules at each location within a sample of biological tissue is confounded by fluorescence produced by the tissue itself (autofluorescence). Spectral mixing models use mixing coefficients to specify how much fluorescence from each source is present and unmixing algorithms separate the two fluorescent sources. Current spectral unmixing methods for AF removal often require a priori knowledge of mixing coefficients. Those which do not, such as principal component analysis, generate negative mixing coefficients that are not physically meaningful. Non-negative matrix factorization constrains mixing coefficients to be non-negative, and has been used for spectral unmixing, but not AF removal. This paper describes a novel non-negative matrix factorization algorithm which separates fluorescent images into true signal and AF components utilizing an estimate of the dark current. We also present a test-bed, based on fluorescent beads, to compare the performance of different AF removal algorithms. Our algorithm out-performed previous state of the art on validation images.

In this paper, the variable view imaging system (VVIS) developed to offer flexibility in microscopic observation is applied to microassembly tasks to overcome several problems in normal imaging systems that hindered application of vision-guided micromanipulation techniques such as limited field of view (FOV) and fixed viewing direction. In three representative cases of microassembly, the parts are mated by visual servoing, avoiding FOV problem by combining unit FOV's and occlusion problem by changing the viewing direction to the optimal one. The results demonstrate the superiority and usefulness of the VVIS in various micromanipulation tasks.

Microassembly has become an important technique to fabricate micro devices with different materials, complex shapes, and incompatible processes. Vision-based microassembly is a promising technique for automated microassembly. However, conventional vision-based techniques in microassembly are limited by inherent problems such as a small depth-of-field (DOF) and a narrow field-of-view (FOV). Microassembly operations initially need to detect micro parts in a wide FOV and a large DOF yet also maintain high resolution for the final state. A tradeoff between the DOF (and/or FOV) and resolution prevents the conventional systems from satisfying the aforementioned microassembly requirements. This paper presents an active zooming control method that enables dynamic adjustment of the DOF (FOV) according to the position and the focus measure of micro objects. The proposed method is based on an artificial potential field method with the capability to combine different kinds of constraints such as the FOV, the focus measure, and the joint limits into the system. The stability and robustness of the proposed system are also investigated. The novelty of this method is that it can ensure the vision system maintains a wide FOV and a large DOF initially, and high resolution at the end. Simulations and microassembly experimental results are provided to verify the feasibility of the proposed approach.

Active shape models (ASMs) are often limited by the inability of relatively few eigenvectors to capture the full range of biological shape variability. This paper presents a method that overcomes this limitation, by using a hierarchical formulation of active shape models, using the wavelet transform. The statistical properties of the wavelet transform of a deformable contour are analyzed via principal component analysis, and used as priors in the contour's deformation. Some of these priors reflect relatively global shape characteristics of the object boundaries, whereas, some of them capture local and high-frequency shape characteristics and, thus, serve as local smoothness constraints. This formulation achieves two objectives. First, it is robust when only a limited number of training samples is available. Second, by using local statistics as smoothness constraints, it eliminates the need for adopting ad hoc physical models, such as elasticity or other smoothness models, which do not necessarily reflect true biological variability. Examples on magnetic resonance images of the corpus callosum and hand contours demonstrate that good and fully automated segmentations can be achieved, even with as few as five training samples.

A method for automated segmentation of major cortical sulci on the outer brain boundary is presented, with emphasis on automatically determining point correspondence and on labeling cortical regions. The method is formulated in a general optimization framework defined on the unit sphere, which serves as parametric domain for convoluted surfaces of spherical topology. A statistical shape model, which includes a network of deformable curves on the unit sphere, seeks geometric features such as high curvature regions and labels such features via a deformation process that is confined within a spherical map of the outer brain boundary. The limitations of the customary spherical coordinate system, which include discontinuities at the poles and nonuniform sampling, are overcome by defining the statistical prior of shape variation in terms of projections of landmark points onto corresponding tangent planes of the sphere. The method is tested against and shown to be as accurate as manually defined segmentations.