π― What it does: A 3D fetal shape and posture statistical model based on SMPL has been constructed, achieving precise alignment and visualization of fetal 3D shapes through alternating optimization of posture and shape on MRI time series; an automated method for conventional fetal measurements is also provided.
CodeClassificationSegmentationDomain AdaptationTransformerVision Language ModelContrastive LearningMultimodalityBiomedical DataMagnetic Resonance Imaging
π― What it does: This paper proposes a new unvalidated, unbalanced support set evaluation framework and a training-free linear probe SS-Text+ for few-shot adaptation in medical visual language models, aimed at achieving robust few-shot adaptation in real medical scenarios.
π― What it does: The FIND-Net framework is proposed, which effectively reduces CT metal artifacts by integrating frequency domain and spatial domain convolutions.
π― What it does: This paper constructs a fine-grained rib fracture diagnosis framework, first using a Faster R-CNN-based detector to locate fractures, and then employing a multi-head classifier and hyperplane multimodal embedding to achieve multi-label classification across four dimensions: location, displacement, morphology, and quantity.
Fine-tuning Vision Language Models with Graph-based Knowledge for Explainable Medical Image Analysis
Li, Chenjun (Cornell University), Paetzold, Johannes C. (Cornell Tech)
CodeClassificationExplainability and InterpretabilityGraph Neural NetworkSupervised Fine-TuningVision Language ModelImageBiomedical Data
π― What it does: This paper proposes a method that combines a heterogeneous graph of biological information with a visual language model, utilizing graph neural networks to stage diabetic retinopathy on OCTA images and generate interpretable diagnostic reports.
π― What it does: Proposes the Flexibly Distilled 3D Rectified Flow (FDRF) framework for predicting brain images and tissue segmentation at 12 or 24 months from 6 months of brain MRI;
π― What it does: A lightweight Flip Distribution Alignment Variational Autoencoder (FDA-VAE) has been designed and implemented to achieve multi-phase CE MRI image synthesis.
π― What it does: A medical image synthesis framework called MOTFM based on optimal transport flow matching is proposed, which can quickly generate high-quality medical images under various modalities, dimensions, and conditions, and can be used for tasks such as generation, segmentation, and denoising.
π― What it does: A framework called FMM-Diff based on diffusion models is proposed, which can generate high-level sequences (such as DWI or T1ce) through feature mapping and fusion modules when multi-modal MRI is missing.
π― What it does: This paper proposes a GLCM-based Masked Autoencoder (GLCM-MAE) pre-training framework that utilizes texture information to enhance the performance of medical image classification models.
FOCUS: Feature Replay with Optimized Channel-Consistent Dropout for U-Net Skip-Connections
Joham, Simon Johannes (Medical University of Graz), Urschler, Martin (Medical University of Graz)
CodeSegmentationDomain AdaptationSafty and PrivacyConvolutional Neural NetworkImageBiomedical DataMagnetic Resonance Imaging
π― What it does: This paper proposes a feature replay framework named FOCUS for domain incremental continuous medical image segmentation, which retains the skip connections of U-Net while meeting privacy and storage constraints.
π― What it does: A multimodal learning framework FMM TC based on foundational models is proposed and implemented to predict drug responses in patients with neuropathic pain.
π― What it does: This paper presents FoundBioNet, a foundational model based on SWIN-UNETR, which non-invasively predicts IDH mutation status in multiparametric MRI by integrating tumor-aware feature encoding and T2-FLAIR differences.
FPN-in-FPN: A Nested Multi-Scale Aggregation Network for Polyp Segmentation
Ye, Jin (Monash University), Cai, Jianfei (Monash University)
CodeSegmentationConvolutional Neural NetworkImage
π― What it does: A nested multi-scale aggregation network (FPN-in-FPN) is designed and implemented for colon polyp segmentation, combining bidirectional feature fusion and deep supervision.
π― What it does: This paper proposes FMISeg, a language-guided medical image segmentation model that achieves fine segmentation through bidirectional interaction of high and low-frequency visual features and text features.
π― What it does: A multi-granularity context network based on frequency domain enhancement, FMC-Net, is proposed for efficiently and accurately segmenting individual vertebrae in 3D CT and MRI images.
π― What it does: FSA-Net is proposed for precise segmentation of the white line (WLT) in laparoscopic images, and the first high-quality LTS (White Line of Toldt Segmentation) dataset is constructed.
FunBench: Benchmarking Fundus Reading Skills of MLLMs
Wei, Qijie (Renmin University of China), Li, Xirong (Renmin University of China)
CodeTransformerLarge Language ModelVision Language ModelImageMultimodalityBenchmark
π― What it does: This paper constructs FunBench, a hierarchical fundus image benchmark based on visual question answering, to systematically evaluate the fundus reading ability of multimodal large language models (MLLMs).
π― What it does: In the field of single-source domain generalization (SDG) for medical image segmentation, the MEDU framework is proposed to enhance the model's generalization performance in scenarios with very few labeled samples.
Future Slot Prediction for Unsupervised Object Discovery in Surgical Video
Liao, Guiqiu (University of Pennsylvania), Hashimoto, Daniel A. (University of Pennsylvania)
CodeObject DetectionSegmentationTransformerVideoBiomedical Data
π― What it does: This paper proposes an unsupervised object discovery framework based on dynamic slot attention and future slot prediction (DTST + slot merging), which can generate interpretable object slots in real-time and reconstruct segmentation masks in surgical videos.
FViM: Frequency Vision Mamba for Label-Free Cell Death Pathway Prediction in Lung Cancer Chemotherapy
Ye, Zhaoyi (Wuhan University), Lei, Cheng (Wuhan University)
CodeClassificationRecognitionExplainability and InterpretabilityDrug DiscoveryImageBiomedical Data
π― What it does: A label-free high-throughput cell death pathway prediction framework based on multidimensional optical time-stretch imaging flow cytometry (OTS-IFC) and frequency vision Mamba (FViM) was developed, applied to the prediction of cell death status and pathways in lung cancer chemotherapy.
π― What it does: The GA-SAM framework is proposed, which utilizes point cloud to generate global 3D shapes with only three slices of sparse annotations, and adapts the Segment Anything Model (SAM) through geometric constraints to achieve precise medical image segmentation.
Tian, Xin (University of Bristol), Zhang, Hang (Cornell University)
CodeOptimizationImageMagnetic Resonance Imaging
π― What it does: This paper proposes an iterative deformation registration framework based on Gaussian primitives (GPO), which achieves precise registration of retinal images by setting control nodes at significant vascular features and using Gaussian-weighted KNN to propagate displacement information.
π― What it does: Designed and implemented a Transformer-based GeneMorphFormer model to predict the two-dimensional coordinates of the gray/white matter boundary curve in the cerebral cortex using gene expression data.
π― What it does: Train a self-supervised vision transformer based on DINOv2 as a foundational model in the field of fetal ultrasound, and evaluate it on classification, segmentation, and few-shot tasks.
π― What it does: MorphLDM is proposed, which synthesizes 3D brain MRI by learning deformable templates and generating deformation fields to capture morphological details.
π― What it does: By using generative unsupervised anomaly detection, a coarse-fine hierarchical ensemble model (CMM and FGM) is employed to identify anomalies in brain CT scans, thereby reducing the workload in emergency radiology.
π― What it does: A framework for detecting anatomical landmarks of anterior teeth in dental CBCT images based on few-shot learning, called GeoSapiens, is proposed to address the issues of scarce labeling and high costs of manual annotation.
Geometry-Guided Local Alignment for Multi-View Visual Language Pre-Training in Mammography
Du, Yuexi (Yale University), Dvornek, Nicha C. (Yale University)
CodeClassificationRecognitionSegmentationTransformerVision Language ModelContrastive LearningImageTextMultimodalityMagnetic Resonance Imaging
π― What it does: Trained a CLIP-based breast imaging-text pre-training model GLAM, utilizing a multi-view geometry-guided local alignment mechanism to learn the correspondence between breast images and reports.
π― What it does: A bone suppression method based on the Global-Local Consistency Model (GL-LCM) is proposed, achieving rapid high-resolution bone suppression in chest X-ray images.
GoCa: Trustworthy Multi-Modal RAG with Explicit Thinking Distillation for Reliable Decision-Making in Med-LVLMs
Dai, Pengyu (Institute of Integrated Research, Institute of Science Tokyo), Suzuki, Kenji (Institute of Integrated Research, Institute of Science Tokyo)
CodeRetrievalExplainability and InterpretabilityKnowledge DistillationTransformerLarge Language ModelContrastive LearningMultimodalityBiomedical DataRetrieval-Augmented GenerationChain-of-Thought
π― What it does: Proposes the GoCa multimodal retrieval-augmented generation (RAG) system, which enhances the credibility and interpretability of Med-LVLM using Chain-of-Thought (CoT) distillation and multi-agent collaboration.
π― What it does: Decomposing the radial and tangential diffusion signals of the cerebral cortex, a GPU-based probabilistic optimization framework is proposed.
GradInvDiff: Stealing Medical Privacy in Federated Learning via Diffusion-Based Gradient Inversion
Wang, Zhiyuan (Beijing Institute of Technology), Liu, Kun (Beijing Institute of Technology)
CodeFederated LearningSafty and PrivacyAdversarial AttackDiffusion modelImageBiomedical DataMagnetic Resonance Imaging
π― What it does: This paper proposes GradInvDiff, a method for implementing gradient inversion attacks on medical images using diffusion models in a federated learning scenario.
Graph Laplacian Transformer with Progressive Sampling for Prostate Cancer Grading
Junayed, Masum Shah (University of Connecticut), Nabavi, Sheida (University of Connecticut)
CodeClassificationTransformerImageBiomedical Data
π― What it does: This paper proposes a prostate cancer grading system based on graph Laplacian attention Transformer and iterative refinement module, which can adaptively filter high-information patches while maintaining spatial consistency.
π― What it does: A neighborhood-aware network based on graph neural networks (GNAN) is proposed, utilizing eye movement data to achieve weakly supervised segmentation of medical images.
π― What it does: For the task of separating pulmonary arteries and veins, the authors propose an end-to-end learning framework based on graph structure called Graph-PAVNet.
π― What it does: This paper proposes a two-stage qMRI reconstruction method called PUQ, which utilizes phase-related uncertainty guidance. It first recovers multi-phase images and estimates the uncertainty of each phase through an iterative network with MC Dropout, and then performs T1/T2 mapping using pixel-level uncertainty as weights during the parameter fitting stage.
π― What it does: This paper proposes using the intermediate features of a pre-trained diffusion model as a similarity measure to guide the deformation registration network for medical images, achieving semantic alignment in the absence of anatomical structures.
HAGE: Hierarchical Alignment Gene-Enhanced Pathology Representation Learning with Spatial Transcriptomics
Dang, Thao M. (University of Texas at Arlington), Huang, Junzhou (University of Texas at Arlington)
CodeRepresentation LearningConvolutional Neural NetworkContrastive LearningImageMultimodalityBiomedical Data
π― What it does: Proposed the HAGE framework, which utilizes gene co-expression embedding and hierarchical alignment to predict spatial transcriptomic expression from histological images.
Hallucination-Aware Multimodal Benchmark for Gastrointestinal Image Analysis with Large Vision-Language Models
Khanal, Bidur (Rochester Institute of Technology), Bhattarai, Binod (University of Aberdeen)
CodeTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelImageTextMultimodalityBenchmark
π― What it does: A multimodal gastrointestinal endoscopy image dataset, Gut-VLM, has been constructed, which includes diagnostic reports generated by VLM (ChatGPT-4 Omni), hallucination sentence labels annotated by medical experts, and corresponding correction texts. Based on this data, hallucination-aware fine-tuning of VLM has been conducted.
Hard Sample Mining-based Tongue Diagnosis for Fatty Liver Disease Severity Classification
Chen, Tao (Eindhoven University of Technology), Liu, Kunhong (Xiamen University)
CodeClassificationMixture of ExpertsImage
π― What it does: A framework for tongue diagnosis based on hard sample mining (HM-TDF) is proposed to accomplish multi-class classification of fatty liver severity, and the Tongue-FLD tongue image dataset is publicly released.
HARM3-Fusion: Hierarchical Attentional Representation Learning of Multi-Modal, Multi-Temporal, and Multi-Sequence Fusion for Pathological Complete Response Prediction of Head and Neck Squamous Cell Carcinoma
π― What it does: A hierarchical attention-based multimodal fusion framework HARM-Fusion is proposed for prognostic prediction of pathological complete response (pCR) in head and neck squamous cell carcinoma (HNSCC) patients after neoadjuvant immunotherapy (NCIT).
Harnessing EHRs for Diffusion-based Anomaly Detection on Chest X-rays
Kim, Harim (Handong Global University), Hong, Charmgil (Handong Global University)
CodeAnomaly DetectionDiffusion modelImageBiomedical DataElectronic Health Records
π― What it does: The study uses diffusion models combined with structured electronic health records (EHR) to achieve unsupervised anomaly detection in chest X-rays.
HARP: Harmonization and Adaptive Refinement of Pseudo-Labels for Cross-Domain Medical Image Segmentation
Liu, Yulong (University of Science and Technology of China), Sun, Mingzhai (University of Science and Technology of China)
CodeSegmentationDomain AdaptationBiomedical Data
π― What it does: This paper proposes the HARP framework for cross-domain semi-supervised medical image segmentation, combining pseudo-label adaptive filtering and inter-domain harmonization modules.
HASD: Hierarchical Adaption for Pathology Slide-Level Domain-Shift
Liu, Jingsong (Technical University of Munich), SchΓΌffler, Peter J. (Technical University of Munich)
CodeDomain AdaptationSupervised Fine-TuningImageBiomedical Data
π― What it does: A Hierarchical Adaption for Pathology Slide-Level Domain-Shift (HASD) framework is proposed to address the domain shift problem in pathology images at the whole slide level.
π― What it does: A heterogeneous mask attention-guided path convolution model (HM-AGPC) has been developed for diagnostic analysis of functional brain networks.
π― What it does: A hierarchical anatomy-aware guided framework based on Transformer is proposed to reconstruct brain tissue microstructure mapping using T1-weighted MRI.
CodeClassificationConvolutional Neural NetworkMixture of ExpertsContrastive LearningImageUltrasound
π― What it does: This study proposes a hierarchical Corpus-View-Category Refinement Framework (CVC-RF) for risk stratification of carotid plaques under multi-view ultrasound.
π― What it does: A hierarchical point cloud feature learning framework based on the state space model (SSM) is proposed for classification, reconstruction, and segmentation tasks of medical point clouds.
π― What it does: This paper proposes a hierarchical, part-based 3D vascular generation framework that separates the modeling of global tree topology from local geometric details, achieving more refined and coherent vascular network synthesis.
Hierarchical Self-Supervised Adversarial Training for Robust Vision Models in Histopathology
Malik, Hashmat Shadab (Mohamed Bin Zayed University of Artificial Intelligence), Khan, Salman (Mohamed Bin Zayed University of Artificial Intelligence)
CodeSegmentationAdversarial AttackConvolutional Neural NetworkGenerative Adversarial NetworkContrastive LearningImageBiomedical Data
π― What it does: This paper proposes Hierarchical Self-Supervised Adversarial Training (HSAT), which generates adversarial samples using the patient-slice-patch three-level relationship and trains the model under a multi-level contrastive learning framework to enhance the robustness of pathological images.
π― What it does: This paper proposes a hierarchical spatio-temporal segmentation network HSS-Net for the segmentation of the left ventricular endocardium in cardiac ultrasound videos, and accurately estimates the ejection fraction based on the segmentation results.
π― What it does: A text-conditioned latent diffusion framework is proposed to achieve one-to-many synthesis of multimodal medical images using a single model.
High-Order Progressive Trajectory Matching for Medical Image Dataset Distillation
Dong, Le, Mou, Lichao (Xidian University)
CodeClassificationData SynthesisKnowledge DistillationConvolutional Neural NetworkImageBiomedical Data
π― What it does: A high-order progressive trajectory matching method is proposed for distilling medical imaging datasets, generating privacy-friendly and efficient synthetic datasets.
HiLa: Hierarchical Vision-Language Collaboration for Cancer Survival Prediction
Cui, Jiaqi (Sichuan University), Wang, Yan (East China Normal University)
CodeClassificationRecognitionTransformerLarge Language ModelVision Language ModelContrastive LearningMultimodalityBiomedical DataMagnetic Resonance Imaging
π― What it does: A model based on hierarchical visual-language collaboration predicts the survival of cancer patients using multi-level visual features and diverse language prompts.
π― What it does: A historical report-guided dual-modal parallel learning framework (BiGen) is proposed, which generates pathology reports for whole slide images (WSI) through visually and textually learnable joint encoding and knowledge retrieval.
π― What it does: A framework for classifying polyps in full-field white light endoscopy is proposed, utilizing alignment-free dense knowledge distillation to transfer diagnostic information from the NBI domain to the WLI domain.
π― What it does: HoloPointNet is proposed, a deep learning framework capable of directly processing 3D point clouds for holographic phase mapping, achieving rapid generation of multi-plane holographic images.
HRVVS: A High-resolution Video Vasculature Segmentation Network via Hierarchical Autoregressive Residual Priors
Yao, Xincheng (Hong Kong University of Science and Technology), Zhu, Lei (Hong Kong University of Science and Technology)
CodeSegmentationTransformerOptical FlowVideo
π― What it does: The HRVVS model is proposed for liver vessel segmentation in high-resolution liver resection surgery videos, and the first Hepa-SEG dataset is constructed.
CodeSegmentationGraph Neural NetworkTransformerImageBiomedical Data
π― What it does: Proposes the Hybrid Graph Mamba (HGM) model for colon polyp segmentation, combining Mamba and GCN for multi-scale feature extraction and fusion.
π― What it does: This paper proposes a hybrid network CRAViM and a cyclic denoising consistency training framework for unpaired medical image synthesis, addressing the issues of global structural distortion and local detail loss caused by traditional convolutional methods.
π― What it does: A hybrid perspective attention network is proposed for the classification of clinically significant prostate cancer in 3D TRUS images.
π― What it does: A hypernetwork-based test-time domain adaptation framework, HyDA, is proposed, which utilizes implicit domain representations to dynamically generate model parameters, enabling the model to adapt to different medical imaging domains without target domain training samples.
HyperPath: Knowledge-Guided Hyperbolic Semantic Hierarchy Modeling for WSI Analysis
Huang, Peixiang (University of Hong Kong), Yu, Lequan (University of Hong Kong)
CodeClassificationVision Language ModelImageBiomedical Data
π― What it does: The HyperPath method is proposed, which utilizes text concept-guided hierarchical representation in hypercurvature space for the classification of whole slide images (WSI).
π― What it does: The HyperSORT framework is proposed, which uses a hypernetwork to predict UNet parameters based on the latent vectors of image and annotation variations, achieving self-organizing robust training.
CodeSegmentationConvolutional Neural NetworkTransformerImageBiomedical Data
π― What it does: A basic model for medical image segmentation based on SAM, called SyncSAM, is proposed, which uses a synchronized dual-branch encoder and a multi-scale dual-branch decoder, significantly improving the segmentation performance of medical images.
π― What it does: Using standardized interactive videos with adult samples, this study extracts and models multimodal non-verbal features such as eye movement, facial expressions, vocal intonation, head movements, and heart rate variability, aiming to achieve computer-aided detection of Autism Spectrum Conditions (ASC).
Improving Motor Imagery EEG Signal Quality with Dynamic Visual Cues: An Innovative Paradigm and Dataset
Yue, Chenxi (Northwestern Polytechnical University), Zhang, Shu (Northwestern Polytechnical University)
CodeClassificationRecognitionConvolutional Neural NetworkTransformerTime SeriesBiomedical Data
π― What it does: A motion imagination EEG acquisition paradigm based on real human motion dynamic visual prompts is proposed, and a corresponding dataset is constructed.
Inferring Super-Resolved Gene Expression by Integrating Histology Images and Spatial Transcriptomics with HISTEX
Xue, Shuailin (Yunnan University), Min, Wenwen (Shenzhen Research Institute of Big Data)
CodeGenerationSuper ResolutionTransformerSupervised Fine-TuningImageMultimodalityBiomedical Data
π― What it does: By integrating histological images with low-resolution spatial transcriptomics data, a super-resolution gene expression map is generated using bidirectional cross-attention and multi-instance learning.
π― What it does: This paper studies how the choice of classification task and the type of distribution shift affect out-of-distribution (OOD) detection and rejection prediction in fetal ultrasound images by comparing the performance of eight uncertainty quantification methods across four different classification tasks.
π― What it does: This paper proposes an Information Bottleneck-based Causal Attention (IBCA) model for multi-label medical image recognition, which effectively separates task-related and unrelated information, thereby improving diagnostic accuracy.
π― What it does: This paper proposes the Instrument-Splatting framework, which combines 3D Gaussian splatting with CAD models and utilizes real monocular surgical videos to achieve controllable and highly realistic digital twin reconstruction of surgical instruments.
π― What it does: A multimodal brain network graph Transformer model (MEGATF) is proposed, which utilizes meta-analysis information for the integration of functional and structural data of brain regions and disease classification.
Intelligent Virtual Sonographer (IVS): Enhancing Physician-Robot-Patient Communication
Song, Tianyu (Technical University of Munich), Navab, Nassir (Technische UniversitΓ€t MΓΌnchen)
CodeSegmentationRobotic IntelligenceTransformerLarge Language ModelImageBiomedical DataUltrasound
π― What it does: Designed and implemented an embedded conversational virtual ultrasound examination assistant IVS based on dual-instance LLM to enhance real-time communication among doctors, robots, and patients, and to achieve real-time command translation and robot control in an XR environment.
π― What it does: An interactive 3D CT image lesion morphology report generation framework is proposed, which combines point annotation for fine segmentation and simultaneously generates a structured report.
Interpretable fMRI Captioning via Contrastive Learning
Shen, Vyacheslav (KAIST), Kim, Daeshik (KAIST)
CodeRetrievalExplainability and InterpretabilityTransformerLarge Language ModelContrastive LearningMultimodalityBiomedical DataMagnetic Resonance Imaging
π― What it does: This study proposes a two-stage contrastive learning framework that utilizes the BLIP-2 model to map fMRI signals into a visual-language embedding space, enabling text descriptions of stimulus images (fMRI captioning) and multimodal retrieval.
π― What it does: This paper proposes using voxel-level brain age prediction as a self-supervised pre-training task to improve brain MRI segmentation models, particularly enhancing performance in low-sample scenarios.
π― What it does: This paper proposes a medical image segmentation network called KMUNet, which integrates KAN and Mamba. It uses a CNN encoder to extract local features in a U-shaped structure and introduces the Mamba module in the decoder to capture global dependencies, while also designing the KAN-PatchEmbed and KAN-SCA modules.
Knowledge-Enhanced Complementary Information Fusion with Temporal Heterogeneous Graph Learning for Disease Prediction
Yang, Zongbao (Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences), Wang, Ruxin (Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences)
CodeGraph Neural NetworkTransformerContrastive LearningGraphTime SeriesBiomedical DataElectronic Health Records
π― What it does: This paper proposes the KCIF framework, which constructs a Temporal Heterogeneous Admission Graph (THAG) and integrates complementary information from laboratory tests and treatment events to achieve accurate predictions of ICU patient conditions.
Learning Contrastive Multimodal Fusion with Improved Modality Dropout for Disease Detection and Prediction
Gu, Yi (OMRON SINIC X Corporation), Ma, Jiaxin (OMRON SINIC X Corporation)
CodeClassificationAnomaly DetectionTransformerContrastive LearningMultimodalityTabularBiomedical DataComputed TomographyElectronic Health Records
π― What it does: This study investigates a framework that combines improved modality loss (learnable modality tokens + simultaneous modality dropout) with multimodal contrastive learning for disease detection and prediction using CT images and electronic health records.
Learning Disease State from Noisy Ordinal Disease Progression Labels
Schmidt, Gustav (University of TΓΌbingen), MΓΌller, Sarah (University of TΓΌbingen)
CodeClassificationConvolutional Neural NetworkImageBiomedical Data
π― What it does: This paper utilizes noisy ordinal disease progression labels (better, stable, worse) to train a Siamese network, learning continuous representations of disease states, and applies this representation for few-shot activity classification on an external OCT dataset.
CodeExplainability and InterpretabilityTransformerBiomedical DataMagnetic Resonance ImagingAlzheimer's Disease
π― What it does: This paper presents NeuroPathX, an interpretable deep learning framework based on early fusion, designed to jointly analyze MRI structural images and genetic polymorphism (SNP) data to explore the interactions between neuroimaging and genetic pathways.
Bassi, Pedro R. A. S. (John Hopkins University), Zhou, Zongwei (University of California, San Francisco)
CodeSegmentationConvolutional Neural NetworkTransformerLarge Language ModelImageTextBiomedical DataComputed Tomography
π― What it does: A new framework called R-Super is proposed, which utilizes radiology reports for direct supervision of CT tumor segmentation. It maps tumor counts, sizes, and locations from the reports to voxel-level supervision through two new loss functions (volume loss and spherical loss) and is trained and evaluated on multiple datasets.
Learning with Explicit Topological Priors for Chest X-ray Rib Segmentation
Zhao, Xiaowei (Anhui University), Li, Chuanfu (Anhui University of Chinese Medicine)
CodeSegmentationConvolutional Neural NetworkImage
π― What it does: In chest X-ray images, a framework is proposed to segment each rib by explicitly embedding the topological priors (connectivity and interaction) of the ribs into the network.
LEAVS: An LLM-based Labeler for Abdominal CT Supervision
Bigolin Lanfredi, Ricardo (National Institutes of Health), Summers, Ronald M. (Icahn School of Medicine at Mount Sinai)
CodeClassificationAnomaly DetectionTransformerLarge Language ModelPrompt EngineeringImageTextComputed TomographyChain-of-Thought
π― What it does: Utilizing a large language model through the zero-shot prompting system LEAVS to extract abnormal labels and urgency levels for various organs from abdominal CT reports, aimed at training imaging models.
π― What it does: Translate CT perfusion images of patients with acute ischemic stroke from CT to MRI, and incorporate lesion-aware image spatial loss during the post-training phase of the latent diffusion model.
CodeClassificationExplainability and InterpretabilityTransformerImageMultimodalityBiomedical DataMagnetic Resonance Imaging
π― What it does: A lesion-centered Vision Transformer (LC-ViT) is proposed, which combines DWI MRI and 62 clinical variables to predict the modified Rankin Scale (mRS) outcomes 3 months after stroke.
Leveraging Diffusion Models for Continual Test-Time Adaptation in Fundus Image Classification
Liu, Mingsi (University of Electronic Science and Technology of China), Xu, Yanwu (South China University of Technology)
CodeClassificationDomain AdaptationDiffusion modelImageBiomedical Data
π― What it does: A continuous testing adaptation framework based on diffusion models, DiffCTA, is proposed for fundus image classification without modifying the source model.
π― What it does: A decentralized brain graph learning framework LG-DBGL based on hemispheric decoupling is proposed for Alzheimer's disease recognition.
π― What it does: A lightweight unsupervised denoising framework called Noise2Detail is proposed, which achieves high-quality denoising on single medical images through a three-stage process.
π― What it does: A method is proposed to transform the problem of reducing contrast agent dosage into a constrained optimization inverse problem by learning a forward operator from high dose to low dose, thereby generating simulated images at high dose levels.
π― What it does: A low-latency organization tracking method called LiteTracker is proposed, based on CoTracker3 and achieving real-time end-to-end tracking.
LKA: Large Kernel Adapter for Enhanced Medical Image Classification
Zhu, Ziquan (University of Leicester), Liu, Zhe (Shenzhen University)
CodeClassificationTransformerImageBiomedical Data
π― What it does: This paper proposes the Large Kernel Adapter (LKA), which enhances the receptive field by introducing channel-level large kernel convolutions when adapting pre-trained models, thereby improving the accuracy of medical image classification.
Localization Lens for Improving Medical Vision-Language Models
Farooq, Hasan (Lahore University of Management Sciences), Mahmood, Arif (Information Technology University of the Punjab)
CodeRecognitionSegmentationTransformerVision Language ModelContrastive LearningImageMultimodalityBiomedical DataMagnetic Resonance Imaging
π― What it does: This paper proposes the 'Positioning Lens' - a set of expert-validated multiple medical image enhancement representations, combined with pixel rearrangement and decoupled contrastive loss, to improve the performance of medical visual-language models (Med-VLM) in anatomical structure and spatial positioning reasoning.
Longitudinal anatomical attention maps for recognizing diagnostic errors from radiologistsβ eye movements
Anikina, Anna (University of Copenhagen), Ibragimov, Bulat (University of Iowa)
CodeRecognitionSegmentationRecurrent Neural NetworkTransformerImageBiomedical Data
π― What it does: This paper proposes a long-term attention mapping method that combines eye-tracking information with visual Transformers to predict diagnostic errors in chest X-ray readings.
LVPNet: A Latent-variable-based Prediction-driven End-to-end Framework for Lossless Compression of Medical Images
Song, Chenyue (Harbin Institute of Technology), Li, Xiang (Nanyang Technological University)
CodeCompressionConvolutional Neural NetworkImageBiomedical Data
π― What it does: An end-to-end lossless medical image compression framework named LVPNet is proposed, which utilizes a Global Multi-Scale Perception Module (GMSM) to extract low-dimensional latent variables and generates pixel distributions through a prediction module, followed by entropy coding.