Zero-shot Referring Image Segmentation Augmented by Multimodal Large Language Models

Pneumothorax requires rapid recognition and accurate interpretation of chest X-rays (CXRs), particularly in acute settings where delays can have serious consequences. With the emergence of advanced image interpretation models capable of visual analysis, their diagnostic reliability in radiology practice remains to be determined. This study aimed to assess the diagnostic performance of three state-of-the-art systems in detecting pneumothorax using a large, well-annotated dataset. A total of 10,675 CXRs from the publicly available SIIM-ACR Pneumothorax Segmentation dataset were analyzed. Three multimodal models (GPT-4o, Gemini 2 Pro, and Claude 4 Sonnet) were evaluated using a uniform, image-based approach. Each model's binary outputs (presence: 1, absence: 0) were compared with reference results to determine accuracy, sensitivity, specificity, precision, and F1 scores. Additional subgroup analyses were conducted across pneumothorax size categories: small, medium, and large. Pairwise statistical comparisons were performed using McNemar's test. Sensitivity, specificity, and overall accuracy are reported with corresponding 95% confidence intervals. The prevalence of pneumothorax in the dataset was 22.3% (n = 2,379). All models demonstrated high specificity (above 0.90) but consistently low sensitivity (0.16-0.36). The best overall performance was observed with Gemini 2, which achieved an accuracy of 0.79 and specificity of 0.95, whereas Claude 4 showed greater sensitivity (0.20-0.34) across lesion-size categories. Diagnostic performance improved with increasing pneumothorax size, but smaller lesions remained difficult to identify. Pairwise comparisons confirmed statistically significant differences among all evaluated systems (P < 0.050). In this large-scale evaluation, the tested models exhibited strong reliability in identifying normal examinations but limited ability to detect subtle or small pneumothoraxes. Despite high specificity, low sensitivity limits the use of current Multimodal large language models as rule-out tools for pneumothorax. With continued refinement, these models may eventually support radiologists by improving workflow efficiency and diagnostic confidence. Automated systems capable of high specificity but low sensitivity should not be relied upon to exclude pneumothorax. However, they may serve as valuable assistants for confirming positive findings and prioritizing urgent cases in busy clinical workflows.

Recently, Referring Image Segmentation (RIS) frameworks that pair the Multimodal Large Language Model (MLLM) with the Segment Anything Model (SAM) have achieved impressive results. However, adapting MLLM to segmentation is computationally intensive, primarily due to visual token redundancy. We observe that traditional patch-wise visual projectors struggle to strike a balance between reducing the number of visual tokens and preserving semantic clarity, often retaining overly long token sequences to avoid performance drops. Inspired by text tokenizers, we propose a novel semantic visual projector that leverages semantic superpixels generated by SAM to identify “visual words” in an image. By compressing and projecting semantic superpixels as visual tokens, our approach adaptively shortens the token sequence according to scene complexity while minimizing semantic loss in compression. To mitigate loss of information, we propose a semantic superpixel positional embedding to strengthen MLLM’s awareness of superpixel geometry and position, alongside a semantic superpixel aggregator to preserve both fine-grained details inside superpixels and global context outside. Experiments show that our method cuts visual tokens by \(\sim\) 93% without compromising performance, notably speeding up MLLM training and inference, and outperforming existing compressive visual projectors on RIS.

Sentiment analysis is crucial in understanding user opinions, emotions, and attitudes across various domains. Traditional sentiment analysis methods rely primarily on textual data, limiting their ability to capture the full context of human expression, often including multimodal elements such as images, audio, and videos. Existing approaches struggle with ambiguity, sarcasm, and lack of contextual awareness, reducing accuracy and effectiveness. To address these limitations, we propose a novel framework called Sentiment Analysis using Machine Learning with Multimodal Models (SA-ML-MM), which integrates text, images, and audio inputs using large language models (LLMs) enhanced with multimodal capabilities. Our approach leverages ML-based feature extraction and fusion techniques to improve sentiment classification accuracy. The proposed framework is applied in social media analysis, customer feedback interpretation, and emotion recognition. Experimental results demonstrate that SA-ML-MM significantly outperforms traditional text-based models, achieving higher accuracy and robustness in sentiment prediction. By incorporating multiple data modalities, our approach effectively captures nuanced emotional expressions, making sentiment analysis more precise and reliable.

Accurate tinnitus diagnosis is crucial for enabling timely therapeutic intervention and longitudinal treatment monitoring. While non-invasive neuroimaging modalities-particularly electroencephalography (EEG) with millisecond temporal resolution and functional magnetic resonance imaging (fMRI) with millimeter spatial resolution- provide complementary neural features, existing diagnostic approaches remain constrained to unimodal analysis of EEG or fMRI data, inherently limiting diagnostic precision and clinical generalizability. This paper introduces TinnitusLLM, the first multimodal large language model (LLM) framework that synergistically integrates EEG and fMRI features for tinnitus diagnosis. To enable LLM-based interpretation of neural signals, this framework integrates three key components: (1) a neuroinspired positional encoding mechanism that injects neurophysiological priors into the embedding space, enabling neurologically grounded, dynamic positional mapping of EEG and fMRI tokens; (2) multimodal autoregressive pretraining on more than 500 hours of EEG and 250 hours of fMRI data to learn causally informed predictive representations; and (3) fine-tuning with a cross-modal, subject-invariant adversarial learning strategy that enforces subject-independent constraints in the shared cross-modal feature space, thereby substantially improving diagnostic robustness across subjects. We validate TinnitusLLM through comprehensive experiments on a rigorously collected multimodal dataset containing 20 participants. Quantitative evaluations demonstrate that TinnitusLLM achieves superior cross-subject diagnostic accuracy compared to the state-of-the-art baseline methods. These results underscore TinnitusLLM's potential as a clinically viable framework for objective tinnitus assessment through multimodal neural decoding.

This study aimed to evaluate the diagnostic accuracy of multimodal large language models in classifying superior labial frenulum attachments from intraoral photographs using expert consensus as the reference standard. Five experts (two periodontists and three orthodontists) established the consensus standard by classifying frenulum attachments in 117 intraoral images as mucosal, gingival, papillary, and papilla penetrating. The same photographs were then presented to three multimodal large language models (ChatGPT 4o, Gemini 2.5 Pro, and Microsoft Copilot GPT-4), and their diagnostic performance was evaluated using accuracy, sensitivity, specificity, and F1 score. Reliability was assessed using Fleiss' and Cohen's Kappa, and diagnostic performances were compared using Cochran's Q test. Human raters demonstrated almost perfect agreement (κ = 0.838, p < 0.001), whereas large language models showed poor inter-model agreement (κ = -0.124, p < 0.001). ChatGPT achieved slight significant agreement with the consensus (κ = 0.114, p = 0.019), although its clinical relevance was negligible. Gemini (κ = 0.099) and Copilot (κ = 0.027) showed no significant agreement (p > 0.05). Copilot yielded the highest overall accuracy (46.2%), followed by Gemini (44.5%) and ChatGPT (35.0%). The performance of the large language models varied across frenulum types. Current multimodal large language models demonstrate inconsistent and clinically insufficient accuracy in the classification of superior labial frenulum attachments from photographs. Domain-specific training is essential before large language models can be considered reliable diagnostic tools in dentistry.

The integration of visual encoders and large language models (LLMs) has driven recent progress in multimodal large language models (MLLMs). However, the scarcity of high-quality instruction-tuning data for vision-language tasks remains a challenge. The current leading paradigm, such as LLaVA, relies on language-only GPT-4 to generate data, which requires pre-annotated image captions and detection bounding boxes, suffering from understanding image details. A practical solution to this problem would be to utilize the available multimodal large language models to generate instruction data for vision-language tasks. However, it's worth noting that the currently accessible MLLMs are not as powerful as their LLM counterparts, as they tend to produce inadequate responses and generate false information. As a solution for addressing the current issue, this paper proposes the Visual Instruction Generation and Correction (VIGC) framework that enables multimodal large language models to generate instruction-tuning data and progressively enhance its quality on-the-fly. Specifically, Visual Instruction Generation (VIG) guides the vision-language model to generate diverse instruction-tuning data. To ensure generation quality, Visual Instruction Correction (VIC) adopts an iterative update mechanism to correct any inaccuracies in data produced by VIG, effectively reducing the risk of hallucination. Leveraging the diverse, high-quality data generated by VIGC, we finetune mainstream models and validate data quality based on various evaluations. Experimental results demonstrate that VIGC not only compensates for the shortcomings of language-only data generation methods, but also effectively enhances the benchmark performance. The models, datasets, and code are available at https://opendatalab.github.io/VIGC

Multimodal Large Language Models (MLLMs) promise seamless integration of vision and language understanding. However, despite their strong performance, recent studies reveal that MLLMs often fail to effectively utilize visual information, frequently relying on textual cues instead. This survey provides a comprehensive analysis of the vision component in MLLMs, covering both application-level and architectural aspects. We investigate critical challenges such as weak spatial reasoning, poor fine-grained visual perception, and suboptimal fusion of visual and textual modalities. Additionally, we explore limitations in current vision encoders, benchmark inconsistencies, and their implications for downstream tasks. By synthesizing recent advancements, we highlight key research opportunities to enhance visual understanding, improve cross-modal alignment, and develop more robust and efficient MLLMs. Our observations emphasize the urgent need to elevate vision to an equal footing with language, paving the path for more reliable and perceptually aware multimodal models.

Local image information is crucial for accurate segmentation of image with intensity inhomogeneity. However, the local information is not embedded in Chan-Vese model. In this paper, we propose a novel level set model which takes the local boundary information into account. The proposed model can overcome the difficulty that CV model suffered, i.e., the unsuccessful segmentation of object with intensity inhomogeneity. Finally, we validate the efficiency of our model on some synthetic and real images.

Multimodal large language models (MLLMs) capable of integrating visual and textual information represent a promising advancement for clinical applications requiring image interpretation. Wound care assessment, which demands simultaneous analysis of wound photographs and clinical data, provides an ideal domain to evaluate multimodal vs unimodal artificial intelligence capabilities against human expertise. This study aims to compare the performance of MLLMs, unimodal ChatGPT 5.0, and human clinical experts on a standardized wound care certification examination. This cross-sectional comparative study evaluated 3 participant groups on a 25-question wound care certification examination spanning 4 clinical domains (Diagnosis, Treatment, Complication Management, and Wound Subtype Knowledge). Participants included 3 MLLMs (Med-PaLM 2, LLaVA-Med, and BioGPT), 1 unimodal large language model (ChatGPT 5.0), and 4 human clinical experts (general surgeon, wound care nurse, and 2 internal medicine physicians). Statistical analyses included one-way ANOVA with Tukey post hoc tests and domain-specific Kruskal-Wallis comparisons. Human experts achieved the highest accuracy (mean 86%, SD 9.1%), followed by MLLMs (mean 78.7%, SD 12.2%), while ChatGPT 5.0 achieved 64% accuracy, failing the 70% certification threshold. Significant overall group differences were observed (F2,5=8.42, P=.02, η²=0.74). MLLMs significantly outperformed ChatGPT 5.0 (difference=14.7 percentage points, P=.03, Cohen d=1.38), with the multimodal advantage most pronounced in visually dependent domains: Diagnosis (81% vs 43%, P=.008) and Complication Management (72% vs 50%, P=.03). No multimodal advantage was observed for text-based Wound Subtype Knowledge (both 67%). Med-PaLM 2 achieved 92% accuracy, matching that of the wound care nurse, while the general surgeon achieved the highest overall performance (96%). MLLMs demonstrate significant performance advantages over unimodal artificial intelligence in wound care assessment, particularly for visually dependent clinical tasks. While human experts with specialized wound care experience maintain overall superiority, the point estimate of the top-performing MLLM (Med-PaLM 2, 92%) fell within the observed range of human scores; however, the underpowered comparison (power=0.52) and wide CIs preclude definitive conclusions regarding noninferiority or equivalence to human experts. These findings support the potential role of MLLMs as clinical decision-support tools, warranting further adequately powered validation studies.

Security concerns related to Large Language Models (LLMs) have been extensively explored; however, the safety implications for Multimodal Large Language Models (MLLMs), particularly in medical contexts (MedMLLMs), remain inadequately addressed. This paper investigates the security vulnerabilities of MedMLLMs, focusing on their deployment in clinical environments where the accuracy and relevance of question-and-answer interactions are crucial for addressing complex medical challenges. We introduce and redefine two attack types: mismatched malicious attack (2M-attack) and optimized mismatched malicious attack (O2M-attack), by integrating existing clinical data with atypical natural phenomena. Using the comprehensive 3MAD dataset that we developed, which spans a diverse range of medical imaging modalities and adverse medical scenarios, we performed an in-depth analysis and proposed the MCM optimization method. This approach significantly improves the attack success rate against MedMLLMs. Our evaluations, which include white-box attacks on LLaVA-Med and transfer (black-box) attacks on four other SOTA models, reveal that even MedMLLMs designed with advanced security mechanisms remain vulnerable to breaches. This study highlights the critical need for robust security measures to enhance the safety and reliability of open-source MedMLLMs, especially in light of the potential impact of jailbreak attacks and other malicious exploits in clinical applications. Warning: Medical jailbreaking may generate content that includes unverified diagnoses and treatment recommendations. Always consult professional medical advice.

Accurate prediction of communication link quality metrics is essential for vehicle-to-infrastructure (V2I) systems, enabling smooth handovers, efficient beam management, and reliable low-latency communication. The increasing availability of sensor data from modern vehicles motivates the use of multimodal large language models (MLLMs) because of their adaptability across tasks and reasoning capabilities. However, MLLMs inherently lack three-dimensional spatial understanding. To overcome this limitation, a lightweight, plug-and-play bird’s-eye view (BEV) injection connector is proposed. In this framework, a BEV of the environment is constructed by collecting sensing data from neighboring vehicles. This BEV representation is then fused with the ego vehicle’s input to provide spatial context for the large language model. To support realistic multimodal learning, a co-simulation environment combining CARLA simulator and MATLAB-based ray tracing is developed to generate RGB, LiDAR, GPS, and wireless signal data across varied scenarios. Instructions and ground-truth responses are programmatically extracted from the ray-tracing outputs. Extensive experiments are conducted across three V2I link prediction tasks: line-of-sight (LoS) versus non-line-of-sight (NLoS) classification, link availability, and blockage prediction. Simulation results show that the proposed BEV injection framework consistently improved performance across all tasks. The results indicate that, compared to an ego-only baseline, the proposed approach improves the macro-average of the accuracy metrics by up to 13.9%. The results also show that this performance gain increases by up to 32.7% under challenging rainy and nighttime conditions, confirming the robustness of the framework in adverse settings.

With advancements in computer technology, the benefits of embodied intelligence are increasingly evident. This interactive learning model allows AI to be more flexibly deployed across diverse fields. Recent developments in multi-modal large language models (LLMs) have accelerated AI progress, especially in autonomous driving. This perspective highlights how embodied intelligence can enhance LLM applications in the mining industry, presenting new opportunities and potential to revolutionize the field. It also examines the challenges of deploying embodied agents in mining and offers insights into future research and development.

In this work, we introduce Mini-Gemini, a simple and effective framework enhancing multi-modality Vision Language Models (VLMs). Despite the advancements in VLMs facilitating basic visual dialog and reasoning, a performance gap persists compared to advanced models like GPT-4 and Gemini. We propose a novel approach to narrow the gap by mining the potential of VLMs for better performance across various cross-modal tasks. It tackles the following questions: (1) How can high-resolution visual tokens improve image understanding without lengthening the token sequence? (2) How to improve reasoning and generation abilities of VLM with high-quality data? (3) How to close the gap between open-source VLMs and proprietary models on reasoning-driven generation? In particular, to enhance visual tokens, we propose to utilize an additional visual encoder for high-resolution refinement without increasing the visual token count. We further construct a high-quality dataset that promotes precise image comprehension and reasoning-based generation, expanding the operational scope of current VLMs. In general, Mini-Gemini further mines the potential of VLMs and empowers current frameworks with image understanding, reasoning, and generation simultaneously. The proposed model supports a series of dense and MoE Large Language Models (LLMs) from 2B to 34B, which achieve leading performance in several zero-shot benchmarks and even surpasses the developed private models. It is demonstrated to attain 80.6% accuracy on the MMB benchmark (+5.4 vs Gemini Pro) and 74.1% on TextVQA (+4.6 vs LLaVA-NeXT), achieving leading performance in several zero-shot benchmarks and even surpasses the developed private models. Furthermore, Mini-Gemini is proven to improve consistently with stronger LLM, visual encoder, and data in experiments.

Rapid development of mobile devices has led to explosive growth of videos and online platforms, which creates great demand for online advertising in videos. Existing advertising methods often aim to randomly select a time point as insertion position, which means that the video content is likely not related to the ad content, resulting in unsatisfactory user experience. While previous works have neglected to understand rich semantics as well as multimodal information in video advertising, in contrast to previous works, we present an innovative method for video-in-video advertising using multimodal modeling. First, different pre-trained models are used to extract multimodal representations. Then, through multimodal modeling, we learn the complementarity among different representations and obtain a unified video-level description. Finally, the unified representations of ads and videos are utilized to find the best matching result for each advertisement. Our method emphasizes the content similarity between ad and video, which would make the transition between video and ad more natural. Comprehensive experiments with both objective and subjective evaluations demonstrate the effectiveness and user-friendliness of our proposed video-in-video advertising framework.

Background: Multimodal large language models (LLMs) are increasingly being explored as clinical support tools, yet their capacity for orthodontic biomechanical reasoning has not been systematically evaluated. This retrospective study assessed their ability to analyze treatment mechanics and explored their potential role in supporting orthodontic decision-making. Methods: Five publicly available models (GPT-o3, Claude 3.7 Sonnet, Gemini 2.5 Pro, GPT-4.0, and Grok) analyzed 56 standardized intraoral photographs illustrating a diverse range of active orthodontic force systems commonly encountered in clinical practice. Three experienced orthodontists independently scored the outputs across four domains—observation, interpretation, biomechanics, and confidence—using a 5-point scale. Inter-rater agreement and consistency were assessed, and statistical comparisons were made between models. Results: GPT-o3 achieved the highest composite score (3.34/5.00; 66.8%), significantly outperforming all other models. The performance ranking was followed by Claude (57.8%), Gemini (52.6%), GPT-4.0 (48.8%), and Grok (38.8%). Inter-rater reliability among the expert evaluators was excellent, with ICC values ranging from 0.786 (Confidence Evaluation) to 0.802 (Observation). Model self-reported confidence showed poor calibration against expert-rated output quality. Conclusions: Multimodal LLMs show emerging potential for assisting orthodontic biomechanical assessment. With expert-guided validation, these models may contribute meaningfully to clinical decision support across diverse biomechanical scenarios encountered in routine orthodontic care.