arXiv subjects

Quoc-Huy Trinh

Publications and source records attributed to Quoc-Huy Trinh.

At least 19 recordsLinked to original sources

arXiv2026

AURA: Unified Multimodal Framework for Conversational Music Editing

Instruction-guided music editors typically process each request independently, limiting their ability to support workflows in which users progressively refine a track. We introduce AURA, a unified multimodal framework for conversational music editing. AURA uses a multimodal large language model to interpret the complete dialogue history, an optional image, and reference audio, distilling the editing intent into compact concept tokens. A concept-to-audio module injects these tokens and frame-aligned reference features into a frozen MusicGen backbone, enabling precise edits while preserving unaffected content. AURA optimizes only 91M parameters while retaining 1.9B frozen backbone parameters. Experiments on Slakh2100 and MoisesDB demonstrate substantial improvements in edit correctness and content preservation over existing instruction-guided methods, including a 4-5 times reduction in FAD for out-of-domain addition and removal.

cs.SD↗

arXiv2026

Disease Burden over Skin Tone: Decomposing the Dermatology-AI Generalization Gap

Dermatology artificial intelligence (AI) models are predominantly trained on light-skinned, cancer-focused image collections, yet they are increasingly proposed for deployment in resource-constrained settings where patients differ from training populations along two confounded axes: skin tone and disease distribution. We investigate whether poor generalization is primarily caused by skin-tone underrepresentation or disease-distribution shift. We evaluate a cancer-trained baseline (ResNet-50 fine-tuned on HAM10000 and ISIC 2019), two dermatology foundation models (DermLIP and MONET), and a general-purpose vision model (DINOv3) as frozen feature extractors. Models are evaluated on a tone-stratified disease-matched dataset (Diverse Dermatology Images, DDI) and a disease-shifted tone-diverse dataset (Skin Condition Image Network, SCIN). Our results show that disease-distribution shift contributes more than skin tone in the evaluated settings. The cancer baseline decreases from 0.62 to 0.21 balanced accuracy when transferred to unfamiliar clinical conditions, while the within-disease skin-tone gap is smaller (0.10-0.18) and inconsistent. Label-free representation analysis shows that this failure reflects a representational limitation rather than only missing output labels: cancer-specialized features poorly cluster unfamiliar conditions (kNN purity lift +0.06 over chance), whereas dermatology-pretrained features retain stronger transferable structure (+0.23). Finally, we show that representation quality predicts recoverable performance under lightweight adaptation. Starting from dermatology foundation models, approximately ten labeled examples per clinical category recover most attainable performance. We release the evaluation protocol and code to support reproducible auditing of dermatology AI generalization.

cs.CV↗

arXiv2026

How Context Attribution Handles What the Model Already Knows

Context attribution methods for large language models (LLMs) identify which input context contributes to the model response. Recent works show the initial success in attributing the con- tributive score of the contexts. However, we observe that when the context overlaps with the training data, these methods can- not disentangle in-context from in-weight (IW) contributions, producing unreliable scores. Based on this observation, in this work, we introduce: 1) an evaluation protocol that relies on four new metrics (base-model context attribution score (BCS), cross-model context attribution consistency (CAC), attribution preservation score (APS), source separation pre- cision (SSP)) and 2) a benchmark dataset (WMDP-Cyber++) with ground-truth provenance labels to systematically assess attribution under IW overlap. In our experiments across four well-known context attribution methods, we demonstrate that they provide unfaithful attribution when the knowledge from the context also exists in the weights. Finally, we adapt these methods for source separation (IW vs. in-context learning (ICL)) and show that they cannot do the disentanglement based on the contributive score

cs.CL↗

arXiv2026

Rad-JEPA 3D: Radiology Joint-Embedding Predictive Model for 3D Computed Tomography

Self-supervised pretraining is central to 3D medical image analysis, where unlabeled CT volumes are abundant but expert annotations are scarce. Yet existing volumetric encoders often fail to preserve the coarse spatial and geometric structure that downstream reasoning depends on, limiting their performance on organ disentanglement, abnormality detection, and spatial understanding when paired with language models. We introduce Rad-JEPA 3D, a joint-embedding predictive framework that learns volumetric CT representations by predicting the latent features of a complete scan from a masked view. At its core is a hybrid H-Mamba encoder that fuses a Mamba state-space branch, which models inter-slice continuity through sequential scanning, with a grouped-query attention branch, which captures cross-plane spatial context, combined through a lightweight per-token router. To improve the quality of intermediate representations, we further propose Hidden States Orthogonal Regularization (HSOR), which aligns student-teacher hidden states and reduces feature redundancy throughout the encoder. This layer-wise regularization produces more consistent and discriminative volumetric representations, leading to improved performance on organ recognition and spatial reasoning tasks. Pretrained on approximately 120,000 CT scans, Rad-JEPA 3D attains state-of-the-art results despite its compact size: with only 4.0B total parameters, it achieves competitive results with state-of-the-art on closed-ended VQA and the best average spatial-reasoning score on the Spatial-Med benchmark. Ablation studies confirm that the hybrid block and HSOR contribute complementary gains, and that the induced spatial structure can substitute for raw language-model scale on volumetric reasoning tasks.

cs.CV↗

arXiv2026

Beyond Medical Diagnostics: How Medical Multimodal Large Language Models Think in Space

Visual spatial intelligence is critical for medical image interpretation, yet remains largely unexplored in Multimodal Large Language Models (MLLMs) for 3D imaging. This gap persists due to a systemic lack of datasets featuring structured 3D spatial annotations beyond basic labels. In this study, we introduce an agentic pipeline that autonomously synthesizes spatial visual question-answering (VQA) data by orchestrating computational tools such as volume estimation and bounding boxes extraction with multi-agent collaboration and expert radiologist validation. We present SpatialMed, the first comprehensive benchmark for evaluating 3D spatial intelligence in medical MLLMs, comprising 31,253 question-answer pairs across multiple organs and tumor types. Our evaluations on 24 state-of-the-art MLLMs and extensive analyses reveal that current models lack robust spatial reasoning capabilities for medical imaging.

cs.CV↗

arXiv2026

SRMA-Mamba: Spatial Reverse Mamba Attention Network for Pathological Liver Segmentation in MRI Volumes

Liver cirrhosis plays a critical role in the prognosis of chronic liver disease. Early detection and timely intervention are essential for reducing mortality rates. However, the intricate anatomical architecture and diverse pathological changes of liver tissue complicate the accurate detection and characterization of pathological liver structures in clinical settings. Existing methods underutilize spatial anatomical details in volumetric MRI data, thereby hindering their clinical effectiveness and explainability. To address this challenge, we introduce a novel Mamba-based network, SRMA-Mamba, designed to model the spatial relationships within complex anatomical structures of MRI volumes. By integrating the Spatial Anatomy-Based Mamba module (SABMamba), SRMA-Mamba performs selective Mamba scans within pathological liver tissues and combines anatomical information from the sagittal, coronal, and axial planes to construct a global spatial context representation, enabling efficient volumetric segmentation of pathological liver structures. Furthermore, we introduce the Spatial Reverse Mamba Attention module (SRMA), designed to progressively refine boundary details in the segmentation map, utilizing both the coarse segmentation map and hierarchical encoding features. Extensive experiments demonstrate that SRMA-Mamba surpasses state-of-the-art methods, delivering exceptional performance in 3D pathological liver segmentation. The source code is available at https://github.com/JunZengz/SRMA-Mamba.

cs.CV↗

arXiv2026

Revisiting LLM Adaptation for 3D CT Report Generation: A Study of Scaling and Diagnostic Priors

Recent advances in multimodal learning, including large language models (LLMs) and vision-language models (VLMs), have demonstrated strong adaptability to natural images. However, extending their use to the medical domain, particularly for volumetric (3D) images, is challenging due to high computational complexity, volumetric dependencies and the semantic gap between visual features and clinical terminology. Naively fine-tuning LLMs on limited medical data often leads to overfitting and clinical hallucination, where linguistic fluency is prioritized over clinical factuality. In this study, we investigate parameter-efficient adaptation strategies for volumetric CT report generation and introduce RAD3D-Prefix, a lightweight diagnostic-prior conditioning framework that minimizes the need for extensive parameter training. This module integrates image embeddings with multi-label diagnostic classification logits, preserving critical clinical details while bridging the semantic gap. By keeping the LLM frozen, our method requires minimal trainable parameters and mitigates the risk of overfitting on small, domain-specific datasets. Through a systematic study spanning LLMs from 96.1M to 1.6B parameters, we find that fine-tuning is most beneficial for smaller LLMs, whereas freezing larger (~1B+ LLMs and training only lightweight projection layers provides a superior trade-off between performance, generalization, and computational efficiency. Across multiple automatic metrics and a clinical reader study, RAD3D-Prefix outperforms comparable parameter-efficient baselines and demonstrates strong out-of-domain generalization while using substantially fewer trainable parameters than fully fine-tuned alternatives.

cs.CL↗

arXiv2026

Firebolt-VL: Efficient Vision-Language Understanding with Cross-Modality Modulation

Recent advances in multimodal large language models (MLLMs) have enabled impressive progress in vision-language understanding, yet their high computational cost limits deployment in resource-constrained scenarios such as personal assistants, document understanding, and smart cameras. Most existing methods rely on Transformer-based cross-attention, whose quadratic complexity hinders efficiency. Moreover, small vision-language models often struggle to precisely capture fine-grained, task-relevant visual regions, leading to degraded performance on fine-grained reasoning tasks that limit their effectiveness in the real world. To address these issues, we introduce Firebolt-VL, an efficient vision-language model that replaces the Transformer-based decoder with a Liquid Foundation Model (LFM) decoder. To further enhance visual grounding, we propose a Token-Grid Correlation Module, which computes lightweight correlations between text tokens and image patches and modulates via the state-space model with FiLM conditioning. This enables the model to selectively emphasize visual regions relevant to the textual prompt while maintaining linear-time inference. Experimental results across multiple benchmarks demonstrate that Firebolt-VL achieves accurate, fine-grained understanding with significantly improved efficiency. Our model and code are available at: https://fireboltvl.github.io

cs.CV↗

arXiv2026

PRS-Med: Position Reasoning Segmentation in Medical Imaging

Prompt-based medical image segmentation has rapidly emerged, yet existing methods rely on explicit prompts like bounding boxes and struggle to reason about the spatial relationships essential for clinical diagnosis. While general-domain models attempt complex coordinate regression, these approaches often lack the structured reliability required for medical applications. In this work, we introduce PRS-Med, a unified framework that adopts an elegant, clinical-first approach to position reasoning segmentation. By utilizing a medical vision-language model integrated with a segmentation decoder, PRS-Med mimics the structured "search patterns" used by radiologists to identify pathologies within specific anatomical zones. To support this robust reasoning, we present the Medical Position Reasoning Segmentation (PosMed) dataset, comprising 116,000 expert-validated, spatially grounded question-answer pairs across six imaging modalities. Unlike previous brittle attempts at spatial reasoning, PosMed leverages a scalable, deterministic pipeline validated by board-certified radiologists to ensure clinical accuracy. Extensive experiments demonstrate that our zone-based reasoning not only improves segmentation accuracy (mean Dice improvements up to +31.2\%) but also provides a high-confidence interpretability layer that outperforms state-of-the-art complex reasoning models. By prioritizing functional reliability over unnecessary technical complexity, PRS-Med offers a practical and scalable baseline for the next generation of intelligent medical assistants.

cs.CV↗

arXiv2025

Viper-F1: Fast and Fine-Grained Multimodal Understanding with Cross-Modal State-Space Modulation

Recent advances in multimodal large language models (MLLMs) have enabled impressive progress in vision-language understanding, yet their high computational cost limits deployment in resource-constrained scenarios such as robotic manipulation, personal assistants, and smart cameras. Most existing methods rely on Transformer-based cross-attention, whose quadratic complexity hinders efficiency. Moreover, small vision-language models often struggle to precisely capture fine-grained, task-relevant visual regions, leading to degraded performance on fine-grained reasoning tasks that limit their effectiveness in the real world. To address these issues, we introduce Viper-F1, a Hybrid State-Space Vision-Language Model that replaces attention with efficient Liquid State-Space Dynamics. To further enhance visual grounding, we propose a Token-Grid Correlation Module, which computes lightweight correlations between text tokens and image patches and modulates the state-space dynamics via FiLM conditioning. This enables the model to selectively emphasize visual regions relevant to the textual prompt while maintaining linear-time inference. Experimental results across multiple benchmarks demonstrate that Viper-F1 achieves accurate, fine-grained understanding with significantly improved efficiency.

cs.CV↗

arXiv2025

NeIn: Telling What You Don't Want

Negation is a fundamental linguistic concept used by humans to convey information that they do not desire. Despite this, minimal research has focused on negation within text-guided image editing. This lack of research means that vision-language models (VLMs) for image editing may struggle to understand negation, implying that they struggle to provide accurate results. One barrier to achieving human-level intelligence is the lack of a standard collection by which research into negation can be evaluated. This paper presents the first large-scale dataset, Negative Instruction (NeIn), for studying negation within instruction-based image editing. Our dataset comprises 366,957 quintuplets, i.e., source image, original caption, selected object, negative sentence, and target image in total, including 342,775 queries for training and 24,182 queries for benchmarking image editing methods. Specifically, we automatically generate NeIn based on a large, existing vision-language dataset, MS-COCO, via two steps: generation and filtering. During the generation phase, we leverage two VLMs, BLIP and InstructPix2Pix (fine-tuned on MagicBrush dataset), to generate NeIn's samples and the negative clauses that expresses the content of the source image. In the subsequent filtering phase, we apply BLIP and LLaVA-NeXT to remove erroneous samples. Additionally, we introduce an evaluation protocol to assess the negation understanding for image editing models. Extensive experiments using our dataset across multiple VLMs for text-guided image editing demonstrate that even recent state-of-the-art VLMs struggle to understand negative queries.

cs.CV↗

arXiv2025

PolypDB: A Curated Multi-Center Dataset for Development of AI Algorithms in Colonoscopy

Colonoscopy is the primary method for examination, detection, and removal of polyps. However, challenges such as variations among the endoscopists' skills, bowel quality preparation, and the complex nature of the large intestine contribute to high polyp miss-rate. These missed polyps can develop into cancer later, underscoring the importance of improving the detection methods. To address this gap of lack of publicly available, multi-center large and diverse datasets for developing automatic methods for polyp detection and segmentation, we introduce PolypDB, a large scale publicly available dataset that contains 3934 still polyp images and their corresponding ground truth from real colonoscopy videos. PolypDB comprises images from five modalities: Blue Light Imaging (BLI), Flexible Imaging Color Enhancement (FICE), Linked Color Imaging (LCI), Narrow Band Imaging (NBI), and White Light Imaging (WLI) from three medical centers in Norway, Sweden, and Vietnam. We provide a benchmark on each modality and center, including federated learning settings using popular segmentation and detection benchmarks. PolypDB is public and can be downloaded at \url{https://osf.io/pr7ms/}. More information about the dataset, segmentation, detection, federated learning benchmark and train-test split can be found at \url{https://github.com/DebeshJha/PolypDB}.

cs.CV↗

arXiv2024

Sing-On-Your-Beat: Simple Text-Controllable Accompaniment Generations

Singing is one of the most cherished forms of human entertainment. However, creating a beautiful song requires an accompaniment that complements the vocals and aligns well with the song instruments and genre. With advancements in deep learning, previous research has focused on generating suitable accompaniments but often lacks precise alignment with the desired instrumentation and genre. To address this, we propose a straightforward method that enables control over the accompaniment through text prompts, allowing the generation of music that complements the vocals and aligns with the song instrumental and genre requirements. Through extensive experiments, we successfully generate 10-second accompaniments using vocal input and text control.

cs.SD↗

arXiv2024

RotCAtt-TransUNet++: Novel Deep Neural Network for Sophisticated Cardiac Segmentation

Cardiovascular disease remains a predominant global health concern, responsible for a significant portion of mortality worldwide. Accurate segmentation of cardiac medical imaging data is pivotal in mitigating fatality rates associated with cardiovascular conditions. However, existing state-of-the-art (SOTA) neural networks, including both CNN-based and Transformer-based approaches, exhibit limitations in practical applicability due to their inability to effectively capture inter-slice connections alongside intra-slice information. This deficiency is particularly evident in datasets featuring intricate, long-range details along the z-axis, such as coronary arteries in axial views. Additionally, SOTA methods fail to differentiate non-cardiac components from myocardium in segmentation, leading to the "spraying" phenomenon. To address these challenges, we present RotCAtt-TransUNet++, a novel architecture tailored for robust segmentation of complex cardiac structures. Our approach emphasizes modeling global contexts by aggregating multiscale features with nested skip connections in the encoder. It integrates transformer layers to capture interactions between patches and employs a rotatory attention mechanism to capture connectivity between multiple slices (inter-slice information). Additionally, a channel-wise cross-attention gate guides the fused multi-scale channel-wise information and features from decoder stages to bridge semantic gaps. Experimental results demonstrate that our proposed model outperforms existing SOTA approaches across four cardiac datasets and one abdominal dataset. Importantly, coronary arteries and myocardium are annotated with near-perfect accuracy during inference. An ablation study shows that the rotatory attention mechanism effectively transforms embedded vectorized patches in the semantic dimensional space, enhancing segmentation accuracy.

cs.CV↗

arXiv2024

SAM-EG: Segment Anything Model with Egde Guidance framework for efficient Polyp Segmentation

Polyp segmentation, a critical concern in medical imaging, has prompted numerous proposed methods aimed at enhancing the quality of segmented masks. While current state-of-the-art techniques produce impressive results, the size and computational cost of these models pose challenges for practical industry applications. Recently, the Segment Anything Model (SAM) has been proposed as a robust foundation model, showing promise for adaptation to medical image segmentation. Inspired by this concept, we propose SAM-EG, a framework that guides small segmentation models for polyp segmentation to address the computation cost challenge. Additionally, in this study, we introduce the Edge Guiding module, which integrates edge information into image features to assist the segmentation model in addressing boundary issues from current segmentation model in this task. Through extensive experiments, our small models showcase their efficacy by achieving competitive results with state-of-the-art methods, offering a promising approach to developing compact models with high accuracy for polyp segmentation and in the broader field of medical imaging.

cs.CV↗

arXiv2024

PGDS: Pose-Guidance Deep Supervision for Mitigating Clothes-Changing in Person Re-Identification

Person Re-Identification (Re-ID) task seeks to enhance the tracking of multiple individuals by surveillance cameras. It supports multimodal tasks, including text-based person retrieval and human matching. One of the most significant challenges faced in Re-ID is clothes-changing, where the same person may appear in different outfits. While previous methods have made notable progress in maintaining clothing data consistency and handling clothing change data, they still rely excessively on clothing information, which can limit performance due to the dynamic nature of human appearances. To mitigate this challenge, we propose the Pose-Guidance Deep Supervision (PGDS), an effective framework for learning pose guidance within the Re-ID task. It consists of three modules: a human encoder, a pose encoder, and a Pose-to-Human Projection module (PHP). Our framework guides the human encoder, i.e., the main re-identification model, with pose information from the pose encoder through multiple layers via the knowledge transfer mechanism from the PHP module, helping the human encoder learn body parts information without increasing computation resources in the inference stage. Through extensive experiments, our method surpasses the performance of current state-of-the-art methods, demonstrating its robustness and effectiveness for real-world applications. Our code is available at https://github.com/huyquoctrinh/PGDS.

cs.CV↗

arXiv2024

Validating polyp and instrument segmentation methods in colonoscopy through Medico 2020 and MedAI 2021 Challenges

Automatic analysis of colonoscopy images has been an active field of research motivated by the importance of early detection of precancerous polyps. However, detecting polyps during the live examination can be challenging due to various factors such as variation of skills and experience among the endoscopists, lack of attentiveness, and fatigue leading to a high polyp miss-rate. Deep learning has emerged as a promising solution to this challenge as it can assist endoscopists in detecting and classifying overlooked polyps and abnormalities in real time. In addition to the algorithm's accuracy, transparency and interpretability are crucial to explaining the whys and hows of the algorithm's prediction. Further, most algorithms are developed in private data, closed source, or proprietary software, and methods lack reproducibility. Therefore, to promote the development of efficient and transparent methods, we have organized the "Medico automatic polyp segmentation (Medico 2020)" and "MedAI: Transparency in Medical Image Segmentation (MedAI 2021)" competitions. We present a comprehensive summary and analyze each contribution, highlight the strength of the best-performing methods, and discuss the possibility of clinical translations of such methods into the clinic. For the transparency task, a multi-disciplinary team, including expert gastroenterologists, accessed each submission and evaluated the team based on open-source practices, failure case analysis, ablation studies, usability and understandability of evaluations to gain a deeper understanding of the models' credibility for clinical deployment. Through the comprehensive analysis of the challenge, we not only highlight the advancements in polyp and surgical instrument segmentation but also encourage qualitative evaluation for building more transparent and understandable AI-based colonoscopy systems.

eess.IV↗

arXiv2024

KDAS: Knowledge Distillation via Attention Supervision Framework for Polyp Segmentation

Polyp segmentation, a contentious issue in medical imaging, has seen numerous proposed methods aimed at improving the quality of segmented masks. While current state-of-the-art techniques yield impressive results, the size and computational cost of these models create challenges for practical industry applications. To address this challenge, we present KDAS, a Knowledge Distillation framework that incorporates attention supervision, and our proposed Symmetrical Guiding Module. This framework is designed to facilitate a compact student model with fewer parameters, allowing it to learn the strengths of the teacher model and mitigate the inconsistency between teacher features and student features, a common challenge in Knowledge Distillation, via the Symmetrical Guiding Module. Through extensive experiments, our compact models demonstrate their strength by achieving competitive results with state-of-the-art methods, offering a promising approach to creating compact models with high accuracy for polyp segmentation and in the medical imaging field. The implementation is available on https://github.com/huyquoctrinh/KDAS.

eess.IV↗