Vision encoders are indispensable for allowing impressive performance of Multimodal Large Language Models (MLLMs) in vision–language tasks such as visual question answering and reasoning. However, existing vision encoders focus on global image representations but overlook fine-grained regional analysis. They are limited in fine-grained perception due to the scarcity of fine-grained annotated data and the lack of a fine-grained pre-training paradigm. In this paper, we propose GranViT, a novel Vision Transformer that integrates fine-grained feature extraction with semantic alignment to Large Language Models (LLMs) via region-level autoregressive training. We first construct Gran-29M, a dataset comprising 29 million natural and OCR images paired with over 180 million high-quality region-level annotations, to enable large-scale fine-grained pretraining. Consequently, we develop a pretraining-adaptation framework along with a self-distillation mechanism to train fine-grained GranViT on Gran-29M. We sufficiently exploit the fine-grained annotations from Gran-29M to resort to bounding-box-to-caption regression to enhance localized visual representation of the vision encoder in the pretraining and caption-to-bounding-box regression to improve vision feature utilization and localization for LLM in the adaptation. We further incorporate a self-distillation mechanism that imposes explicit localization constraints on the vision encoder to strengthen its regional reasoning capability. Extensive experiments show that GranViT surpasses existing vision encoders and attains strong transferability to varying LLMs. Remarkably, it achieves state-of-the-art results on fine-grained recognition, multimodal VQA, and OCR understanding.

Vision–language pre-training (VLP) holds great promise for general-purpose medical AI by leveraging radiology reports as rich textual supervision, yet existing methods struggle with 3D CT imaging due to inefficient visual backbones and coarse semantic alignment. To address these issues, we propose a tailored VLP framework featuring three key components: (1) a CNN–ViT hybrid encoder that replaces ViT’s patch embedding with a 3D CNN backbone to efficiently capture local anatomical details while preserving global attention and compatibility with pre-trained cross-modal priors; (2) a disease-level contrastive learning mechanism using learnable query tokens to dynamically extract disease-specific semantics from full reports and align them with corresponding visual features, thereby disentangling distinct diseases within the same anatomical region; and (3) a diagnosis-aware prompt strategy that employs real clinical phrases and aggregated disease prototypes to bridge the pre-training–inference gap and enhance zero-shot diagnostic reliability. Our model achieves state-of-the-art performance on CT-RATE (84.4\% AUC, +5.1%) and Rad-ChestCT (75.4\% AUC, +5.4%), with even larger gains (+9.8% AUC) on a challenging 60-disease benchmark, and demonstrates strong transferability to radiology report generation, underscoring the generality and clinical utility of our approach.

Vision encoders serve as the cornerstone of multimodal understanding. Single-encoder architectures like CLIP exhibit inherent constraints in generalizing across diverse multimodal tasks, while recent multi-encoder fusion methods introduce prohibitive computational overhead to achieve superior performance using complementary visual representations from multiple vision encoders. To address this, we propose a progressive pruning framework, namely Multi-Encoder collaboraTivE tOken pRuning (METEOR), that eliminates redundant visual tokens across the encoding, fusion, and decoding stages for multi-encoder MLLMs. For multi-vision encoding, we discard redundant tokens within each encoder via a rank guided collaborative token assignment strategy. Subsequently, for multi-vision fusion, we combine the visual features from different encoders while reducing cross-encoder redundancy with cooperative pruning. Finally, we propose an adaptive token pruning method in the LLM decoding stage to further discard irrelevant tokens based on the text prompts with dynamically adjusting pruning ratios for specific task demands. To our best knowledge, this is the first successful attempt that achieves an efficient multi-encoder based vision language model with multi-stage pruning strategies. Extensive experiments on 11 benchmarks demonstrate the effectiveness of our proposed approach. Compared with EAGLE, a typical multi-encoder MLLMs, METEOR reduces 76

Trapped by the label scarcity in molecular property prediction and drug design, graph contrastive learning (GCL) came forward. Leading contrastive learning works show two kinds of view generators, that is, random or learnable data corruption and domain knowledge incorporation. While effective, the two ways also lead to molecular semantics altering and limited generalization capability, respectively. To this end, we relate the LinE graph with MOlecular graph coNtrastive learning and propose a novel method termed LEMON. Specifically, by contrasting the given graph with the corresponding line graph, the graph encoder can freely encode the molecular semantics without omission. Furthermore, we present a new patch with edge attribute fusion and two local contrastive losses enhance information transmission and tackle hard negative samples. Compared with state-of-the-art (SOTA) methods for view generation, superior performance on molecular property prediction suggests the effectiveness of our proposed framework.

Exploring navigation strategies in lunar environment contributes to understanding the unique navigation mechanism of humans in extraterrestrial environments. However, it is unclear whether human navigation strategies in lunar environments are the same as those in common environments. In this study, a virtual lunar exploration navigation experiment was conducted. Participants were required to complete spatial learning, navigation, and destination-pointing tasks while their behavioral performance and scalp electroencephalogram (EEG) data were recorded. The navigation trials (88 trials from 62 participants) were divided into two groups - path retracing strategy (N = 60, navigating along the known routes) and path integration strategy (N = 28, inferring potential shortcuts) groups - and differences in navigation performance and brain workload between them were measured. Results indicated that trials using the path integration strategy were more efficient in terms of time cost and pointing error. Particularly, navigators using the path integration strategy were adaptive in their brain workload. Their EEG theta power spectral density (PSD) metrics differed for routes with different difficulties; this difference was not found in the path retracing group. This study offers insights into human navigation strategies and cognitive processes in virtual lunar scenes and contributes to future human adaptation to the lunar surface environment when conducting space missions.

Sulfur-containing small molecules, mainly including cysteine (Cys), homocysteine (Hcy), glutathione (GSH), and hydrogen sulfide (H2S), are crucial biomarkers, and their levels in different body locations (living cells, tissues, blood, urine, saliva, etc.) are inconsistent and constantly changing. Therefore, it is highly meaningful and challenging to synchronously and accurately detect them in complex multi-component samples without mutual interference. In this work, we propose a steric hindrance-regulated probe, NBD-2FDCI, with single excitation dual emissions to achieve self-adaptive detection of four analytes. This probe was meticulously designed and constructed from a pKa-tuned 2FDCI fluorophore and a thiol-specific recognition moiety NBD. Except for 661 nm fluorescence for indicating the total biothiols and H2S, Cys and Hcy could trigger an additional 550 nm fluorescence. Utilizing the distinctive responses, the probe NBD-2FDCI exhibited exclusive linear ranges for GSH, Cys/Hcy, and H2S to avoid high-level component interference. Thus, the probe was then applied for sulfur compound measurements in urine samples, indicating metabolic disorder of Cys and H2S in bladder cancer patients. Moreover, adaptive imaging of probe NBD-2FDCI in cells was performed with the results being consistent with in vitro testing. In a word, spatial hindrance strategy-guided probes may exhibit broader prospects in the detection of similar components in complex samples.

Existing pretraining methods for semantic segmentation are hampered by the task gap between global image-level pretraining and local pixel-level finetuning. Joint dense-level pretraining is a promising alternative to exploit off-the-shelf annotations from diverse segmentation datasets but suffers from low-quality class embeddings and inconsistent data and supervision signals across multiple datasets by directly employing CLIP. To overcome these challenges, we propose a novel M ulti-datas E t harmo N ized pretraining framework for S emantic s E gmentation (MENSA). MENSA incorporates high-quality language embeddings and momentum-updated visual embeddings to effectively model the class relationships in the embedding space and thereby provide reliable supervision information for each category. To further adapt to multiple datasets, we achieve one-to-many pixel-embedding pairing with cross-dataset multi-label mapping through cross-modal information exchange to mitigate inconsistent supervision signals and introduce region-level and pixel-level cross-dataset mixing for varying data distribution. Experimental results demonstrate that MENSA is a powerful foundation segmentation model that consistently outperforms popular supervised or unsupervised ImageNet pretrained models for various benchmarks under standard fine-tuning. Furthermore, MENSA is shown to significantly benefit frozen-backbone fine-tuning and zero-shot learning by endowing pixel-level distinctiveness to learned representations.

We introduce the EEGET-RSOD, a simultaneous electroencephalography (EEG) and eye-tracking dataset for remote sensing object detection. This dataset contains EEG and eye-tracking data when 38 remote sensing experts located specific objects in 1,000 remote sensing images within a limited time frame. This task reflects the typical cognitive processes associated with human visual search and object identification in remote sensing imagery. To our knowledge, EEGET-RSOD is the first publicly available dataset to offer synchronized eye-tracking and EEG data for remote sensing images. This dataset will not only advance the study of human visual cognition in real-world environment, but also bridge the gap between human cognition and artificial intelligence, enhancing the interpretability and reliability of AI models in geospatial applications.

Background:Although neoadjuvant chemoimmunotherapy (NACI) is increasingly applied in clinical settings, its therapeutic efficacy and prognostic significance remain unclear. This study sought to establish a surface-enhanced Raman spectroscopy (SERS)-based approach for assessing treatment efficacy and predicting prognosis in patients with locally advanced gastric cancer (LAGC) undergoing NACI. In addition, the utility of SERS for molecular and pathological profiling was investigated. Methods:This retrospective study enrolled 31 patients with LAGC treated with anti-PD-1 inhibitors plus chemotherapy before gastrectomy (May 2018-December 2022). A Raman score (RS) was established from SERS spectral features to predict overall survival (OS). The area under the time-dependent receiver operating characteristic curve (AUC), Cox proportional hazards regression, and concordance index (C-index) were used to evaluate model performance. A nomogram combining RS and ypTNM stage was constructed. Kaplan-Meier analysis assessed the risk stratification capacity. Key spectral bands were analyzed for biomarker identification, and machine learning (ML) models were used for histopathological and molecular classification. Results:A total of 3,670 spectra from 31 patients were analyzed. The RS, based on Raman spectral features, achieved AUCs of 0.854 (1-year OS) and 0.920 (3-year OS). Lower RS correlated with longer OS (p<0.05). RS served as an independent prognostic factor in multivariable analysis. The nomogram incorporating RS and ypTNM improved prediction for 3-year OS (AUC = 0.955) while maintaining 1-year accuracy. Kaplan-Meier analysis confirmed effective risk stratification (P = 0.01). Nine significant Raman bands were linked to nucleotides, collagen, and proteins. ML models achieved >0.85 accuracy in classifying microsatellite instability (MSI), combined positive score (CPS) of programmed cell death ligand-1 (PD-L1), and tumor regression grade (TRG) based on SERS data. Conclusions:This study demonstrates that label-free SERS can effectively predict prognosis in NACI-treated LAGC patients and shows promise in molecular and pathological profiling, supporting its potential for clinical application.

Objective The Acute Disease Quality Initiative (ADQI) working group recently released a consensus definition of sepsis-associated acute kidney injury (SA-AKI), but the prognosis and risk factors for early and late SA-AKI have not been studied.Methods This was a retrospective cohort study based on the Medical Information Mart for Intensive Care IV (MIMIC-IV) database (v2.2). First, SA-AKI patients that met the new definition from the ADQI were screened, and then, the relationships between SA-AKI occurrence time and relevant clinical parameters were analyzed.Results After propensity score matching, 1,090 early SA-AKI (AKI occurring within 48 h of sepsis diagnosis) cases and late SA-AKI (AKI occurring between 48 h and 7 days after sepsis diagnosis) cases were identified. Compared with late SA-AKI patients, early SA-AKI patients had no significant differences in all-cause mortality at 28 days, 60 days or 180 days, renal replacement therapy (RRT) rates; or major adverse kidney events at 30 days (MAKE-30). However, the renal recovery of early SA-AKI patients was significantly better than that of late SA-AKI patients, their lengths of hospital stay and intensive care unit stay were significantly shorter, and the number of patients with positive fluid balance was lower, but the use of nonsteroidal anti-inflammatory drugs (NSAIDs) and nephrotoxic antibiotics and the incidence of septic shock were higher. In addition, there was a difference in the number of patients with early and late SA-AKI at highest AKI stages 1 and 3. Data analysis also revealed that liver disease, cancer, highest AKI stage 3 and septic shock were associated with renal nonrecovery.Conclusions Although there was no significant difference in mortality between early and late SA-AKI patients, there were significant differences in renal recovery, positive fluid balance, nephrotoxic antibiotic use, septic shock and AKI stage. Therefore, the classification of early and late SA-AKI has certain scientific and rational validity, but whether the two have different clinical outcomes and pathogeneses requires further study.

Accurate determination of microsatellite instability (MSI) status is critical for tailoring treatment approaches for gastric cancer patients. Existing clinical techniques for MSI diagnosis are plagued by problems of suboptimal time efficiency, high cost, and burdensome experimental requirements. Here, we for the first time establish the classification model of gastric cancer MSI status based on Raman spectroscopy. To begin with, we reveal that tumor heterogeneity-induced signal variations pose a prominent impact on MSI classification. To eliminate this issue, we develop Euclidean distance-based Raman Spectroscopy (EDRS) algorithm, which establishes a standard spectrum to represent the "most microsatellite stable" status. The similarity between each spectrum of tissues with the standard spectrum is calculated to provide a direct assessment on the MSI status. Compared to machine learning-algorithms including k-Nearest Neighbors, Random Forest, and Extreme Learning Machine, the EDRS method shows the highest accuracy of 94.6 %. Finally, we integrate the EDRS method with the clinical diagnostic modality, computed tomography, to construct an innovative joint classification model with good classification performance (AUC = 0.914, Accuracy = 94.6 %). Our work demonstrates a robust, rapid, non-invasive, and convenient tool in identifying the MSI status, and opens new avenues for Raman techniques to fit into existing clinical workflow.

BACKGROUND:Peritoneal metastasis (PM) is the most common form of distant metastasis in gastric cancer and is a major cause of mortality. Current diagnostic approaches suffer from low sensitivity, time-consuming procedures, and cannot provide real-time diagnostic information. Surface-enhanced Raman spectroscopy (SERS) combined with machine learning algorithms has emerged as a promising tool for cancer diagnosis. PATIENTS AND METHODS:Raman spectra were collected from the peritoneal lavage fluid (PLF) of 120 patients with gastric cancer and analyzed using three machine learning models: principal component analysis-linear discriminant analysis (PCA-LDA), random forest (RF), and support vector machine (SVM). The sensitivity, specificity, accuracy, false positive rate, false negative rate, positive predictive value, and negative predictive value were calculated. Receiver operating characteristic curve analysis was used to assess the diagnostic performance. RESULTS:The accuracy, sensitivity, and specificity of SERS analysis to determine PM with PCA-LDA were 95.7%, 87.0%, and 95.5%; with RF were 95.4%, 91.3%, and 96.0%; with SVM were 95.5%, 91.3%, and 96.0%. For exfoliative cytology, these parameters were 72.0%, 40.0%, and 100%. For computed tomography (CT) scan, these parameters were 72.5%, 57.9%, and 85.7%. In addition, the performance of these models (PCA-LDA, RF, and SVM) demonstrated high diagnostic accuracy, with area under the curve values of 96.9%, 92.1%, and 93.4%, respectively. The diagnostic performance of all models in diagnosing PM is significantly better than those of exfoliative cytology and CT imaging. CONCLUSIONS:The integration of SERS with machine learning models provides a simple, convenient, and cost-effective tool for PLF, offering significant potential for improving the diagnosis of PM.

Existing self-supervised methods for gaze estimation using the dominant streams of contrastive and generative approaches are restricted to eye images and could fail in general full-face settings. In this paper, we reveal that contrastive methods are ineffective in data augmentation for self-supervised full-face gaze estimation, while generative methods are prone to trivial solutions due to the absence of explicit regularization on semantic representations. To address this challenge, we propose a novel approach calledBootstrap auto-encoders withContrastive paradigm (BeCa), which combines the strengths of both generative and contrastive methods. Specifically, we revisit the Auto-Encoder used in generative approaches and incorporate the contrastive paradigm to introduce explicit regularization on gaze representation. Furthermore, we design the InfoMSE loss as an alternative to the vanilla MSE loss for Auto-Encoder to mitigate the inconsistency between reconstruction and representation learning. Experimental results demonstrate that the proposed approaches outperform state-of-the-art unsupervised gaze approaches on extensive datasets (including wild scenes) under both within-dataset and cross-dataset protocols.

Vision-language foundation models, represented by Contrastive Language-Image Pre-training (CLIP), have gained increasing attention for jointly understanding both vision and textual tasks. However, existing approaches primarily focus on training models to match global image representations with textual descriptions, thereby overlooking the critical alignment between local regions and corresponding text tokens. This paper extends CLIP with multi-granularity alignment. Notably, we deliberately construct a new dataset comprising pseudo annotations at various levels of granularities, encompassing image-level, region-level as well as pixel-level captions and tags. Accordingly, we develop a Unified Multi-Granularity learning framework, termed UMG-CLIP, which simultaneously empowers the model with versatile perception abilities across different levels of detail. With parameter efficient tuning, UMG-CLIP surpasses current widely used CLIP variants and achieves state-of-the-art performance on diverse image understanding benchmarks, including open-world recognition, retrieval, semantic segmentation, and panoptic segmentation tasks. We believe that UMG-CLIP represents a valuable advancement in vision-language foundation models. The code is available at https://github.com/lygsbw/UMG-CLIP.

Cytotoxicity assays are crucial for assessing the efficacy of drugs in killing cancer cells and determining their potential therapeutic value. Measurement of the effect of drug concentration, which is an influence factor on cytotoxicity, is of great importance. This paper proposes a cytotoxicity assay using microwave sensors in an end-point approach based on the detection of the number of live cells for the first time. In contrast to optical methods like fluorescent labeling, this research uses a resonator-type microwave biosensor to evaluate the effects of drug concentrations on cytotoxicity by monitoring electrical parameter changes due to varying cell densities. Initially, the feasibility of treating cells with ultrapure water for cell counting by a microwave biosensor is confirmed. Subsequently, inhibition curves generated by both the CCK-8 method and the new microwave biosensor for various drug concentrations were compared and found to be congruent. This agreement supports the potential of microwave-based methods to quantify cell growth inhibition by drug concentrations.

Hepatocellular carcinoma (HCC) is characterized by frequent intrahepatic and distant metastases, resulting in a poor prognosis for patients. Epithelial-mesenchymal transition (EMT) plays a pivotal role in this process. However, the expression of NOP2/Sun RNA methyltransferase 5 (NSUN5) in HCC and its role in mediating EMT remain poorly understood. In this study, clinicopathological analyses are conducted across multiple independent HCC cohorts and induced tumor formation in Nsun5-knockout mice. The findings reveal an upregulation of NSUN5 expression in tumor tissues; conversely, the absence of Nsun5 hinders the malignant progression of HCC, indicating that NSUN5 may serve as a significant oncogene in HCC. Furthermore, elevated levels of NSUN5 enhance EMT processes within HCC cells. NSUN5-knockout cells exhibit reduced invasion and migration capabilities under both in vivo and in vitro conditions, while overexpression of NSUN5 yields opposing effects. Mechanistically, high levels of NSUN5 promote the enrichment of trimethylated histone H3 at lysine 4 (H3K4me3) at the promoter region of SMAD3 through recruitment of the WDR5, thereby facilitating HCC metastasis via SMAD3-mediated EMT pathways. Collectively, this study identifies NSUN5 as a novel driver of metastasis in HCC and provides a theoretical foundation for potential therapeutic strategies against this malignancy.

Pre-training followed by full fine-tuning has gradually been substituted by Parameter-Efficient Tuning (PET) in the field of computer vision tasks. PET has gained popularity, especially in the context of large-scale models, due to its ability to reduce transfer learning costs and conserve hardware resources. However, existing PET approaches primarily focus on recognition tasks and typically support uni-modal optimization, neglecting dense prediction tasks and vision language interactions. To address this limitation, we propose a novel PET framework called Bi-directional Intertwined Vision Language Efficient Tuning for Referring Image Segmentation (BarLeRIa), which leverages bi-directional intertwined vision language adapters to fully exploit the frozen pre-trained models' potential in cross-modal dense prediction tasks. In BarLeRIa, two different tuning modules are employed for efficient global and local attention, as well as an intertwined vision language tuning algorithm for efficient modal fusion. Extensive experiments conducted on challenging RefCOCO-related benchmarks demonstrating the superiority of BarLeRIa over prior PET methods with a significant margin, \emph{i.e.}, achieving an average improvement of 5.6\%. Remarkably, without requiring additional training datasets, BarLeRIa even surpasses SOTA full fine-tuning approaches.

Interstitial cystitis/bladder pain syndrome is a chronic pain syndrome of elusive etiology, accompanied by lower urinary tract symptoms. Over the past decades, many studies have been carried out for exploration of more effective therapies against IC/BPS. However, the results have been inconsistent, probably due to the multifactorial nature of IC/BPS. We establish a model of IC/BPS in mice by combining protamine sulfate /lipopolysaccharide and phenylephrine. Typical histological changes and symptoms were observed. We then explored the effectiveness of artesunate (ART), which has been reported to alleviate autoimmune diseases. Phenotypic tests demonstrated a significant reduction in symptoms. Histological staining showed pathological improvement. WGCNA identified three gene modules specifically related to IC/BPS, and six genes were identified as hub genes. CIBERSORT analysis showed that the activated NK cells seem to be decreased in IC modeling group and partially restored in IC + ART group, whereas the resting NK cells showed the opposite trend. Single-cell transcriptomic analysis elaborated on the changing trends of subgroups of infiltrated immune cells, including T cells, NK cells, and dendritic cells. Our study represents our effort in establishing a reliable and reproducible IC/BPS murine model, and the first study using scRNA-seq in exploring the immune microenvironment of the IC/BPS murine model, and the possible molecular mechanisms of ART treatment in IC/BPS. Further studies are needed to confirm the effect of ART in IC/BPS patients.

Since different quantities of white blood cells (WBCs) in solution possess an adaptive osmotic pressure of cells, the WBCs themselves and in solution have similar concentrations, resulting in them having similar dielectric properties. Therefore, a microwave sensor could have difficulty in sensing the quantity variation when WBCs are in solution. This paper presents a highly sensitive, linear permittivity-inspired microwave biosensor for WBCs, counting through the evaporation method. Such a measurement method is proposed to record measurements after the cell solution is dripped onto the chip and is completely evaporated naturally. The proposed biosensor consists of an air-bridged asymmetric differential inductor and a centrally located circular fork-finger capacitor fabricated on a GaAs substrate using integrated passive fabrication technology. It is optimized to feature a larger sensitive area and improved Q-factor, which increases the effective area of interaction between cells and the electromagnetic field and facilitates the detection of their changes in number. The sensing relies on the dielectric properties of the cells and the change in the dielectric constant for different concentrations, and the change in resonance properties, which mainly represents the frequency shift, corresponds to the macroscopic change in the concentration of the cells. The microwave biosensors are used to measure biological samples with concentrations ranging from 0.25 × 106 to 8 × 106 cells per mL in a temperature (26.00 ± 0.40 °C) and humidity (54.40 ± 3.90 RH%) environment. The measurement results show a high sensitivity of 25.06 Hz/cells·mL−1 with a highly linear response of r2 = 0.99748. In addition, a mathematical modeling of individual cells in suspension is performed to estimate the dielectric constant of individual cells and further explain the working mechanism of the proposed microwave biosensor.