🎯 What it does: This paper proposes a semi-supervised medical image segmentation method called MHL3, based on Mutual Mask Mix and consistency of high and low-level features.
🎯 What it does: This paper proposes a few-shot medical anomaly detection framework called MadCLIP based on CLIP, which learns normal and abnormal features through a dual-branch adapter and learnable text prompts, achieving image-level anomaly classification and pixel-level anomaly segmentation.
MAGNET-AD: Multitask Spatiotemporal GNN for Interpretable Prediction of PACC and Conversion Time in Preclinical Alzheimer
Hassan, Salma (Mohamed Bin Zayed University of Artificial Intelligence), Yaqub, Mohammad (Mohamed bin Zayed University of Artificial Intelligence)
CodeExplainability and InterpretabilityDrug DiscoveryGraph Neural NetworkMultimodalityBiomedical DataAlzheimer's DiseaseElectronic Health Records
🎯 What it does: A multi-task spatiotemporal graph neural network called MAGNET-AD is proposed for predicting PACC scores and conversion times in preclinical Alzheimer's disease patients.
🎯 What it does: This paper proposes a complete automated pipeline (MAGO-SP) for detecting and correcting water-fat exchange errors in volumetric images (VIBE) caused by missing magnitude information. The pipeline includes: ① Using the nnU-Net segmentation network to identify whether water or fat regions have been exchanged; ② Employing the Palette Conditional Denoising Diffusion Network to generate a prior for the 'water' image; ③ Using the prior as an initial value, combined with the physical constraints of the MAGO/MAGORINO optimization model to restore accurate water-fat separation.
CodeClassificationRetrievalTransformerLarge Language ModelVision Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: A multi-faceted knowledge-enhanced visual-language pre-training framework called MAKE is proposed, specifically designed for dermatological zero-shot diagnosis, concept annotation, and cross-modal retrieval.
MAMBA-Based Weakly Supervised Medical Image Segmentation with Cross-Modal Textual Information
Pan, Zhen (Shandong Normal University), Zheng, Yuanjie (Shandong Normal University)
CodeSegmentationTransformerContrastive LearningImageTextMultimodalityBiomedical Data
🎯 What it does: A weakly supervised medical image segmentation framework based on Mamba, TIFCMamba, is proposed, utilizing text descriptions as supervisory information and achieving segmentation through cross-modal alignment.
🎯 What it does: A multi-stage segmentation framework MARSeg based on the Masked Autoregressive model is proposed, combined with an adaptive feature fusion module to achieve fine medical image segmentation.
🎯 What it does: A Mask Contrastive Language-Image Modeling (MCLIM) framework is proposed, which combines text descriptions generated from brain atlases with a masked image recovery task to achieve self-supervised pre-training for brain MRI.
🎯 What it does: This paper proposes a framework (MatchGen) for optimizing pseudo-normal images during the testing phase, which improves the accuracy of medical anomaly region detection by minimizing the pixel-level differences between the input image and the generated image.
🎯 What it does: This paper proposes a training-free multi-center adaptive uncertainty-aware prompting strategy (MAUP) that utilizes the pre-trained Segment Anything Model to achieve cross-domain few-shot medical image segmentation.
🎯 What it does: Maverick is proposed, a collaboration-free federated unlearning framework that allows individual clients to locally implement forgetting of their data without the participation of other clients.
🎯 What it does: This paper proposes a multi-dose adaptive attention diffusion model (MDAA-Diff) based on CT guidance, aimed at restoring low-dose PET to standard-dose PET.
🎯 What it does: A multi-domain diffusion prior-guided MRI reconstruction framework (MDPG) is proposed, which combines the prior guidance of a pre-trained latent diffusion model (LDM) in both latent and image domains. By utilizing the Visual-Mamba backbone, Latent Guided Attention (LGA), Dual-Domain Fusion Branch (DFB), and k-space regularization based on the NACS set, the quality of compressed sensing MRI reconstruction is significantly improved.
🎯 What it does: Using a small number (about 30) of medical image-mask pairs, we fine-tuned Stable Diffusion 1.5 to achieve controllable generation of medical image-mask pairs with structural dependencies and domain specificity. We enhanced the generation quality through automatic quality assessment and mask erosion; the generated data was used to augment five medical segmentation datasets, significantly improving the performance of downstream segmentation models.
🎯 What it does: A general category discovery framework for medical images, MedGCD, is proposed, which can cluster new categories and classify known categories by utilizing weak-strong data augmentation, dual strong view consistency, and confidence-aware pairing objectives in the presence of both labeled and unknown categories.
🎯 What it does: A framework called MeDi (Metadata-Guided Diffusion) is proposed and implemented, which simultaneously considers class labels and multidimensional metadata (such as medical center, race, gender, etc.) in the generation of pathological images, thereby achieving targeted data augmentation for underrepresented subgroups and mitigating bias.
Medical Contrastive Learning of Positive and Negative Mentions
Wu, WeiLong (Tsinghua University), Wu, Ji (Tsinghua University)
CodeClassificationRetrievalLarge Language ModelContrastive LearningImageTextMultimodalityBiomedical DataComputed Tomography
🎯 What it does: A visual entailment-based triplet contrastive learning framework VECL is proposed, which achieves more precise cross-modal feature alignment using positive and negative medical report mentions.
CodeClassificationAnomaly DetectionTransformerPrompt EngineeringMixture of ExpertsImageBiomedical DataUltrasound
🎯 What it does: A case-level prenatal abdominal ultrasound anomaly classification method based on multi-instance learning is proposed, which can complete diagnosis without standard plane localization.
MedNNS: Supernet-based Medical Task-Adaptive Neural Network Search
Mecharbat, Lotfi Abdelkrim (Mohammed Bin Zayed University of Artificial Intelligence), Yaqub, Mohammad (Mohamed bin Zayed University of Artificial Intelligence)
CodeClassificationNeural Architecture SearchImageMultimodalityBiomedical Data
🎯 What it does: This paper proposes a medical task adaptive neural network search framework based on a super network, MedNNS, which can simultaneously select the optimal network structure and corresponding pre-trained weights when given a medical dataset.
MedVLM-R1: Incentivizing Medical Reasoning Capability of Vision-Language Models (VLMs) via Reinforcement Learning
Pan, Jiazhen (Technical University of Munich), Rueckert, Daniel (Technical University of Munich)
CodeOptimizationTransformerReinforcement LearningVision Language ModelMultimodalityBiomedical DataMagnetic Resonance ImagingComputed Tomography
🎯 What it does: Designed and trained MedVLM-R1, a visual language model capable of explicitly generating reasoning processes in medical visual question answering, utilizing GRPO reinforcement learning to encourage the model to produce logical reasoning on its own.
MELON: Multimodal Mixture-of-Experts with Spectral-Temporal Fusion for Long-Term MObility EstimatioN in Critical Care
Zhang, Jiaqing (University of Florida), Rashidi, Parisa (University of Florida)
CodeClassificationRecognitionTransformerMixture of ExpertsMultimodalityTime SeriesBiomedical DataElectronic Health Records
🎯 What it does: This paper proposes a dual-branch multimodal framework named MELON for predicting the movement status of ICU patients over a 12-hour period.
Memory-Augmented Incomplete Multimodal Survival Prediction via Cross-Slide and Gene-Attentive Hypergraph Learning
Qu, Mingcheng (Harbin Institute of Technology), Fan, Lei (University of New South Wales)
CodeClassificationGraph Neural NetworkContrastive LearningMultimodalityBiomedical Data
🎯 What it does: A memory-enhanced multimodal survival prediction framework M²Surv is proposed, which can simultaneously handle multiple pathological slides (including FFPE and FF) and genomic data, while maintaining high predictive performance even in the presence of missing modalities.
Memory-Augmented SAM2 for Training-Free Surgical Video Segmentation
Yin, Ming (University of Exeter), Fu, Zeyu (University of Exeter)
CodeSegmentationVideo
🎯 What it does: This paper proposes a training-agnostic MA-SAM2 framework that achieves multi-object surgical video segmentation with a single prompt, addressing the misjudgment and tracking failure issues of SAM2 in fast motion, occlusion, and complex interactions.
🎯 What it does: A multi-task graph Transformer network based on meta-analysis is proposed, which can simultaneously predict neurological diseases (such as schizophrenia and ADHD) and related cognitive deficits, revealing changes in functional connectivity.
🎯 What it does: A SPECT attenuation correction method based on diffusion models has been developed, which achieves decoupling of CT priors through a morphological structure attention fusion module and completes the correction using only SPECT images during the inference phase.
MiCo: Multiple Instance Learning with Context-Aware Clustering for Whole Slide Image Analysis
Li, Junjian (Central South University), Wang, Jianxin (Institute of Guizhou Aerospace Measuring and Testing Technology)
CodeClassificationSegmentationImageBiomedical Data
🎯 What it does: MiCo is proposed under the multi-instance learning framework, utilizing context-aware clustering to improve cross-region semantic association modeling for whole slide images (WSI).
MicroMIL: Graph-Based Multiple Instance Learning for Context-Aware Diagnosis with Microscopic Images
Kim, JongWoo (Korea Advanced Institute of Science and Technology), Yi, Mun Yong (Korea Advanced Institute of Science and Technology)
CodeClassificationObject DetectionGraph Neural NetworkImageBiomedical Data
🎯 What it does: This paper designs and implements a weakly supervised multi-instance learning framework called MicroMIL for traditional optical microscope images, achieving tumor diagnosis through representative image extraction and graph convolutional networks.
🎯 What it does: A post-processing method called Semantically Diverse Reconstructions (SDR) is proposed, which generates a set of semantically diverse reconstructions consistent with the sampled data based on existing MRI reconstructions, in order to reduce the risk of lesion miss detection caused by under-sampling.
🎯 What it does: This paper proposes a feature style mixing method based on regularized flow (MixStyleFlow) to improve the generalization performance of medical image segmentation models on unseen domains.
🎯 What it does: This paper presents MM-DINOv2, a framework that adapts the pre-trained visual foundation model DINOv2 to multimodal medical imaging tasks, achieving robustness to missing sequences and the ability to utilize unlabeled data through multimodal patch embedding, full-modality masking, and semi-supervised learning.
MOC: Meta-Optimized Classifier for Few-Shot Whole Slide Image Classification
Xiang, Tianqi (Hong Kong University of Science and Technology), Li, Xiaomeng (Southern Medical University)
CodeClassificationMeta LearningVision Language ModelImage
🎯 What it does: This paper proposes a Meta-Optimized Classifier (MOC) for few-shot whole slide image classification, which dynamically combines multiple base classifiers through a meta-learner to achieve global optimal discrimination of images.
🎯 What it does: A framework called M2M-Reg is proposed to transform multimodal image registration into unimodal registration, achieving unsupervised and semi-supervised registration through a cyclic structure and gradient cyclic consistency regularization.
More performant and scalable: Rethinking contrastive vision-language pre-training of radiology in the LLM era
Li, Yingtai (University of Science and Technology of China), Zhou, S. Kevin (Hohai University)
CodeClassificationRetrievalConvolutional Neural NetworkLarge Language ModelSupervised Fine-TuningContrastive LearningImageTextMultimodalityComputed Tomography
🎯 What it does: Utilize large language models to automatically extract diagnostic labels from radiology reports, and use this for supervised pre-training on a large scale dataset, followed by aligning CT images with reports using CLIP.
MOSCARD - Multimodal Opportunistic Screening for Cardiovascular Adverse events with Causal Reasoning and De-confounding
Pi, Jialu (Arizona State University), Banerjee, Imon (Arizona State University)
CodeClassificationAnomaly DetectionTransformerContrastive LearningMultimodalityBiomedical DataElectronic Health RecordsElectrocardiogram
🎯 What it does: This paper proposes MOSCARD, a multimodal causal inference model that combines chest X-rays and 12-lead ECGs for opportunistic screening of future MACE risk, enhancing prediction fairness and generalization through debiasing and causal intervention.
MoST-IG: Morphology-Guided Spatial Transcriptomics Integration via Visual-Genomic Graph Optimal Transport
Yu, Liting (University of Hong Kong), Yu, Lequan (University of Hong Kong)
CodeClassificationSegmentationOptimizationGraph Neural NetworkContrastive LearningMultimodalityBiomedical Data
🎯 What it does: A multi-modal framework MoST-IG is proposed, which achieves high-quality fusion across multiple slices by utilizing organizational morphology information and spatial transcriptomics data.
🎯 What it does: This study investigates a framework named MOST, which implements sequential fine-tuning of a single MR reconstruction network across multiple downstream tasks, avoiding catastrophic forgetting.
Motion-Boundary-Driven Unsupervised Surgical Instrument Segmentation in Low-Quality Optical Flow
Liu, Yang (King's College London), Ourselin, Sebastien (King's College London)
CodeObject DetectionSegmentationOptical FlowVideoBiomedical Data
🎯 What it does: A completely unsupervised surgical instrument segmentation method is proposed, utilizing motion boundaries to guide and improve optical flow quality;
MOTOR: Multimodal Optimal Transport via Grounded Retrieval in Medical Visual Question Answering
Shaaban, Mai A. (Mohamed bin Zayed University of Artificial Intelligence), Yaqub, Mohammad (Mohamed bin Zayed University of Artificial Intelligence)
CodeRetrievalOptimizationVision Language ModelImageTextMultimodalityBiomedical DataElectronic Health RecordsRetrieval-Augmented Generation
🎯 What it does: The MOTOR method is proposed, which combines multimodal optimal transport (OT) with image-based and text-based retrieval for re-ranking, thereby improving the accuracy of answers in medical visual question answering (MedVQA).
CodeClassificationAnomaly DetectionMixture of ExpertsVideoBiomedical DataUltrasound
🎯 What it does: An automatic mitral regurgitation diagnosis model MReg based on four-chamber color Doppler ultrasound video has been developed, achieving graded regression prediction for normal, mild, and moderate to severe regurgitation.
🎯 What it does: The first 3D multi-class segmentation dataset for seven types of lesions in the whole abdomen, MSWAL, is proposed, and based on this dataset, an Inception nnU-Net that integrates the Inception module is designed to achieve fine segmentation of gallstones, kidney stones, liver tumors, kidney tumors, pancreatic cancer, liver cysts, and kidney cysts in abdominal CT images.
🎯 What it does: A semi-supervised learning framework MTCNet is proposed, based on motion and topological consistency guidance, for precise segmentation of the mitral valve in 4D ultrasound cardiac images.
Multi-Agent Collaboration for Integrating Echocardiography Expertise in Multi-Modal Large Language Models
Qin, Yi (Hong Kong University of Science and Technology), Li, Xiaomeng (Southern Medical University)
CodeTransformerLarge Language ModelAgentic AIPrompt EngineeringImageTextMultimodalityUltrasound
🎯 What it does: A comprehensive academic knowledge base for echocardiography (ECED) has been constructed, containing over 100 types of heart diseases, charts, texts, and other multimodal information. Efficient knowledge extraction is achieved through the multi-agent collaborative MMLLM tool (MACEE). Furthermore, a lightweight expert-enhanced visual instruction tuning framework (EEVIT) is proposed, which injects ECED knowledge into pre-trained multimodal large language models via the Expert-Lens adapter.
Multi-expert collaboration and knowledge enhancement network for multimodal emotion recognition
Wang, Kun (Nanjing University of Aeronautics and Astronautics), Zhang, Daoqiang (Nanjing University of Aeronautics and Astronautics)
CodeRecognitionTransformerMultimodalityBiomedical Data
🎯 What it does: This paper proposes a multi-expert collaboration and knowledge-enhanced network for fusing EEG and facial expressions for multimodal emotion recognition.
Multi-Level Gated U-Net for Denoising TMR Sensor-Based MCG Signals
Xing, Zeyu (Zhejiang University), Jiang, Tianzi (University of Chinese Academy of Sciences)
CodeRestorationConvolutional Neural NetworkTime SeriesBiomedical Data
🎯 What it does: A multi-layer gated U-Net (MGU-Net) model is proposed and implemented to denoise magnetocardiogram (MCG) signals collected using tunnel magnetoresistance (TMR) sensors.
🎯 What it does: This study investigates a multi-scale attention-based multi-instance learning framework for weakly supervised classification and localization of breast cancer, capable of extracting multi-scale features and dynamically fusing information from different scales under single-scale input.
🎯 What it does: A multi-sensory cognitive computing framework mCOCO based on stochastic Reservoir Computing is proposed, which learns population-level connectivity templates from BOLD signals and introduces multimodal inputs to enhance cognitive abilities.
🎯 What it does: Construct a rs-fMRI brain network and propose a multi-view graph contrastive learning framework MGCL-DA, utilizing dynamic adaptation and cross-sample topology enhancement for automatic diagnosis of brain diseases.
🎯 What it does: Aiming at the classification of primary tumors in multi-source metastatic cervical lymphoma (CLA), a multimodal fusion network (MDFN) is proposed, which integrates B-mode ultrasound (BUS), color Doppler flow imaging (CDFI), and distribution-based tumor marker (TM) imputation.
🎯 What it does: A multimodal hypergraph-guided learning framework is proposed for non-invasive prediction of survival time in patients with ccRCC (clear cell renal cell carcinoma).
🎯 What it does: This paper proposes a multi-stage alignment and fusion framework that utilizes multimodal imaging (T1-MRI, Tau-PET, FOD) and clinical tables (age, gender, MoCA score) to achieve a three-class diagnosis of Alzheimer's disease (CN, MCI, AD).
🎯 What it does: A multi-sequence MRI unsupervised anomaly detection framework called MultiTransAD is proposed, which directly quantifies anomalies using cross-sequence translation errors.
🎯 What it does: A multi-view information bottleneck framework MvHo-IB is constructed, which jointly learns the pairwise connections and triple interaction information of brain regions, and compresses irrelevant redundancy through the information bottleneck.
NAVIUS: Navigated Augmented Reality Visualization for Ureteroscopic Surgery
Acar, Ayberk (Vanderbilt University), Wu, Jie Ying (Florida International University)
CodeImageBiomedical DataComputed Tomography
🎯 What it does: The NAVIUS system is proposed and implemented, which displays a 3D model of the renal pelvis system generated from preoperative CT in real-time on HoloLens 2, along with the EM tracking position of the ureteroscope, helping surgeons to explore the renal pelvis more comprehensively.
CodeClassificationDomain AdaptationAuto EncoderImageBiomedical Data
🎯 What it does: An unsupervised domain-invariant chest X-ray diagnosis framework based on Neighborhood Consistent Binarization Transformation (NCBT) and Intermediate Domain Style Preserving Autoencoder (IDSP-AE) is proposed;
🎯 What it does: This paper addresses the issue of local-global training mismatch between the hash encoder and MLP in NeRF-based CBCT reconstruction, proposing two solutions: normalized hash encoder and mapping consistency initialization, which significantly enhance training stability and convergence speed.
NERO: Explainable Out-of-Distribution Detection with Neuron-level Relevance in Gastrointestinal Imaging
Chhetri, Anju (NepAl Applied Mathematics and Informatics Institute for research), Bhattarai, Binod (University of Aberdeen)
CodeAnomaly DetectionExplainability and InterpretabilityConvolutional Neural NetworkTransformerImageBiomedical Data
🎯 What it does: A post-hoc OOD detection method called NERO is proposed, which determines whether a sample is OOD by clustering neuron-level correlation centers and calculating correlation distances, while also providing interpretability.
🎯 What it does: A neural network framework called Neural Proteomics Fields (NPF) for super-resolution prediction of spatial proteomics (seq-SP) is proposed, and a Pseudo-Visium SP benchmark dataset is constructed.
NeuroXVocal: Detection and Explanation of Alzheimer’s Disease through Non-invasive Analysis of Picture-prompted Speech
Ntampakis, Nikolaos (International Hellenic University), Argyriou, Vasileios (Kingston University London)
CodeClassificationExplainability and InterpretabilityTransformerTextMultimodalityAlzheimer's DiseaseRetrieval-Augmented GenerationAudio
🎯 What it does: Developed NeuroXVocal, an end-to-end interpretable Alzheimer's disease detection system that integrates multimodal audio, text, and speech embeddings.
🎯 What it does: A similarity measure for multimodal image registration based on local functional dependence is proposed, and an efficient implementation with a closed-form solution is achieved through learning a linear basis function model.
🎯 What it does: A multi-scale fusion framework based on calibrated patch weights (CIB) is proposed for 3D segmentation of new multiple sclerosis lesions.
🎯 What it does: Developed the NIMOSEF framework, which jointly achieves high-resolution segmentation, image reconstruction, and displacement field estimation for cardiac MRI.
🎯 What it does: Proposes the NMSW framework, which utilizes differentiable Top-K sampling to select only a small number of high-importance patches during inference, combining low-resolution global predictions and patch-level predictions to complete 3D medical image segmentation, eliminating the high time and memory consumption issues of traditional sliding window inference.
🎯 What it does: This paper proposes a noise-controllable complex-valued diffusion model (NC-CDM) for directly generating k-space data of hyperpolarized 129Xe lung MRI to alleviate the issue of sample scarcity.
🎯 What it does: This paper presents Nora, a noise-robust fine-tuning framework for ultrasound image segmentation, which achieves cross-domain generalization by injecting adaptive noise into the feature space of SAM and constructing instance-aware prompts.
🎯 What it does: A volume label refinement framework VDDR based on discrete diffusion processes is proposed, combined with a Dilating & Selecting mechanism, to refine noise annotations for three types of middle ear ossicles.
🎯 What it does: A multi-granularity mask autoencoder (MG-MAE) is proposed for the segmentation of non-significant objects in medical images, achieving multi-scale feature learning through global pixel reconstruction and local HOG feature refinement.
🎯 What it does: An online medical image segmentation domain adaptation framework ODES is proposed, which combines expert-guided active learning to achieve single forward and backward updates.
OFF-CLIP: Improving Normal Detection Confidence in Radiology CLIP with Simple Off-Diagonal Term Auto-Adjustment
Park, Junhyun (DGIST), Hwang, Minho (DGIST)
CodeClassificationAnomaly DetectionTransformerPrompt EngineeringContrastive LearningImageTextBiomedical DataElectronic Health Records
🎯 What it does: This paper proposes OFF-CLIP, an improved method to enhance the accuracy of normal case detection in radiology CLIP models by incorporating off-diagonal loss and text filtering.
OpenPath: Open-Set Active Learning for Pathology Image Classification via Pre-trained Vision-Language Models
Zhong, Lanfeng (University of Electronic Science and Technology of China), Wang, Guotai (University Of Electronic Science And Technology Of China)
CodeClassificationTransformerLarge Language ModelVision Language ModelImageBiomedical Data
🎯 What it does: This paper proposes OpenPath, an open-set active learning framework for pathological image classification, addressing the cold start and OOD data selection issues.
CodeGenerationData SynthesisSafty and PrivacyTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelVideoText
🎯 What it does: This work presents Ophora, a text instruction-based model for generating ophthalmic surgery videos. It first constructs 160K high-quality video-instruction pairs through a complete data cleaning process and then progressively fine-tunes a pre-trained T2V model using this data, ultimately achieving real-time generation of realistic and reliable surgical videos based on doctor descriptions.
🎯 What it does: This paper proposes a one-shot video segmentation network named OralSAM, which can achieve segmentation of periodontal ultrasound videos using single-frame annotations.
CodeOptimizationExplainability and InterpretabilityBiomedical Data
🎯 What it does: A framework called OTSurv based on multi-instance learning is proposed, which models the global long-tail heterogeneity and local predictive uncertainty of WSI using optimal transport to achieve survival time prediction.
Out-of-Distribution Nuclei Segmentation in Histology Imaging via Liquid Neural Networks with Modern Hopfield Layer
Swain, Bishal Ranjan (Kumoh National Institute of Technology), Ko, Jaepil (Kumoh National Institute of Technology)
CodeSegmentationDomain AdaptationImageBiomedical Data
🎯 What it does: A framework that combines liquid neural networks with modern Hopfield networks is proposed for handling discrete distribution kernel segmentation of omics images.
🎯 What it does: A parameter-efficient fine-tuning method based on tensor networks, TenVOO, is proposed for the 3D U-Net structure of DDPM to generate brain MRI images.
🎯 What it does: Using path signature (PS) features to perform macrostructural reconstruction and change detection of white matter in clinical-grade diffusion MRI data of patients with depression before and after SSRI treatment, revealing the reorganization of tracts such as TPT, ALIC, and SCC.
Pathology Report Generation and Multimodal Representation Learning for Cutaneous Melanocytic Lesions
Lucassen, Ruben T. (University Medical Center Utrecht), Veta, Mitko (Eindhoven University of Technology)
CodeGenerationRetrievalRepresentation LearningTransformerLarge Language ModelVision Language ModelContrastive LearningImageTextMultimodalityBiomedical Data
🎯 What it does: A visual-language model specifically designed for skin melanocytic lesions was trained and evaluated, capable of generating complete pathology reports based on whole-slide H&E images and achieving cross-modal retrieval.
🎯 What it does: Designed and implemented the MedSign framework, embedding watermarks in text-driven medical image generation while preserving pathological information; dynamically adjusting watermark strength by locating pathological regions through cross-attention.
Pathology-Informed Latent Diffusion Model for Anomaly Detection in Lymph Node Metastasis
Wang, Jiamu (Korea University), Kwak, Jin Tae (Korea University)
CodeAnomaly DetectionVision Language ModelDiffusion modelImageBiomedical Data
🎯 What it does: A latent diffusion model based on pathological text prompts (AnoPILaD) is proposed for unsupervised detection of lymph node metastatic abnormalities.
🎯 What it does: The PathoPainter framework is proposed, which generates high-quality, aligned digital pathology image-mask pairs through an image inpainting method based on tumor masks and regional tumor embeddings.
PathVG: A New Benchmark and Dataset for Pathology Visual Grounding
Zhong, Chunlin (Huazhong University of Science and Technology), Bai, Xiang (Huazhong University of Science and Technology)
CodeObject DetectionSegmentationConvolutional Neural NetworkTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelImageTextBiomedical DataBenchmark
🎯 What it does: This paper proposes the PathVG pathological visual localization benchmark and constructs the RefPath dataset. It then designs the PKNet model to enhance pathological text understanding through LLM, completing text-based localization of pathological image regions.
🎯 What it does: This paper proposes a patient-specific feature selection framework that combines generative models and radiomics for disease diagnosis in knee MRI.
🎯 What it does: A self-supervised TOF-PET reconstruction framework is proposed, utilizing implicit neural representation (INR) to model PET images, and achieving reconstruction through differential forward projection and beam-level TV regularization.
CodeImage TranslationGenerationPrompt EngineeringGenerative Adversarial NetworkContrastive LearningImageBiomedical Data
🎯 What it does: A Prompt-Driven Universal Model (PD-UniST) is proposed, capable of converting H&E stained images into various IHC stained images without paired training, supporting the generation of staining results for different tumor markers with a single training session.
🎯 What it does: This paper presents a large-scale panoramic dental X-ray dataset, PerioXrays, and proposes a clinically-oriented framework for detecting apical periodontitis, PerioDet, which combines a background denoising attention module (BDA) and an IoU dynamic calibration module (IDC) to achieve automatic detection of apical periodontitis.
🎯 What it does: A tool segmentation framework called ToolSeg is proposed, which is oriented towards surgical phase information, and the first pixel-level annotated dataset Sankara-MSICS is provided based on MSICS surgery.
🎯 What it does: A two-stage cardiac MRI synthesis framework (CPGG) is proposed, which first generates the cardiac phenotype distribution and then generates high-fidelity cardiac CINE sequences conditionally based on the phenotype using a masked autoregressive diffusion model.
🎯 What it does: A physics-informed implicit neural representation (PINR) framework is proposed for the joint estimation of the B0 field and the reconstruction of distortion-free EPI images.