🎯 What it does: A replay-free class incremental learning framework CRCL is proposed, utilizing a pre-trained visual Transformer to achieve continuous diagnosis of medical images through two types of learners (conservative and aggressive).
ConStyX: Content Style Augmentation for Generalizable Medical Image Segmentation
Chen, Xi (Chinese Academy of Sciences), Zaiane, Osmar R. (Amii)
CodeSegmentationDomain AdaptationConvolutional Neural NetworkImageBiomedical Data
🎯 What it does: The ConStyX method is proposed, which enhances the domain generalization ability of medical image segmentation models by simultaneously performing content and style augmentation in the deep feature space.
🎯 What it does: A method based on a conditional diffusion model for contrast flow pattern and cross-phase specificity perception is proposed, enabling the generation of multi-phase enhanced CT from non-contrast CT.
🎯 What it does: This study proposes an unsupervised domain generalization method for harmonizing MRI scanners based on a diffusion autoencoder, which can harmonize brain MRI images from any unknown scanner to a target domain without requiring multiple scans of the same subject or multiple sequences, while maintaining anatomical structures.
Contrastive Masked Video Modeling for Coronary Angiography Diagnosis
Shao, Zhiming (Chinese Academy of Sciences), Hui, Hui (Chinese Academy of Sciences)
CodeClassificationRecognitionRepresentation LearningTransformerContrastive LearningVideoBiomedical Data
🎯 What it does: A self-supervised framework combining masked video modeling and contrastive learning is proposed for the diagnosis of coronary angiography (CAG).
🎯 What it does: A controllable flow matching model (CFM) is proposed to achieve one-step synthesis from non-contrast T1w brain MRI to contrast T1ce brain MRI, with precise control over tumor details through tumor segmentation constraints.
🎯 What it does: A controllable image synthesis workflow is proposed, utilizing adaptive cell segmentation and style transfer to generate cervical cell images with boundary annotations, enhancing the training of detection models.
🎯 What it does: Generate synthetic cardiac LGE MRI images that meet clinical attributes or segmentation mask conditions using a controllable latent diffusion model (Stable Diffusion), and evaluate the performance of myocardial infarction segmentation methods with these synthetic images.
🎯 What it does: A controllable skin lesion image synthesis model based on Lesion-Focused Vector Autoregression (LF-VAR) is proposed, which can generate high-quality and controllable skin images according to lesion measurement scores and lesion types.
CoPA: Hierarchical Concept Prompting and Aggregating Network for Explainable Diagnosis
Dong, Yiheng (Huazhong University of Science and Technology), Yang, Xin (Shenzhen University)
CodeClassificationExplainability and InterpretabilityTransformerVision Language ModelContrastive LearningImageBiomedical Data
🎯 What it does: The CoPA framework is proposed, which achieves interpretable disease diagnosis through a multi-layer visual concept encoder, a concept alignment bottleneck layer, and a gated aggregation module. This framework utilizes the Concept-aware Embedding Generator (CEG) to extract multi-scale concept representations from each layer and guides the model's focus on specific concepts through Concept Prompt Tuning (CPT), ultimately completing disease prediction through alignment and aggregation.
Core-Periphery Principle Guided State Space Model for Functional Connectome Classification
Chen, Minheng (University of Texas at Arlington), Zhu, Dajiang (University of Georgia)
CodeClassificationMixture of ExpertsBiomedical DataMagnetic Resonance ImagingAlzheimer's Disease
🎯 What it does: A state space model based on the core-edge principle (CP-SSM) has been designed and implemented for binary classification of functional connectivity graphs.
🎯 What it does: CortexGen is proposed, a two-stage generative framework based on geometric variational autoencoders and flow matching, designed to synthesize high-resolution realistic cortical surfaces without the need for MRI data.
🎯 What it does: This paper aligns the emotional semantic features extracted from electroencephalogram (EEG) to electrocardiogram (ECG) through cross-modal contrastive learning, enabling emotion recognition using only ECG.
🎯 What it does: A cross-modal knowledge distillation method is proposed, which transforms 2D chest X-ray (CXR) embeddings into 3D computed tomography pulmonary angiography (CTPA) embeddings through a latent diffusion prior, enabling pulmonary embolism (PE) diagnosis using only CXR.
🎯 What it does: A large-scale multimodal dataset for immunotherapy in NSCLC was constructed, and a cross-modal mask learning framework was proposed for survival prediction.
🎯 What it does: A cross-view general diffusion model (CvGDiff) is proposed for sparse view CT reconstruction, utilizing deterministic artifacts generated by angle downsampling as degradation operators, and sharing semantic information across multiple views.
🎯 What it does: A CSAL-3D framework is proposed to achieve one-shot active learning for 3D medical image segmentation, combining self-supervised learning-driven uncertainty-enhanced diversity sampling.
🎯 What it does: The study focuses on recognizing action triplets (instrument, verb, target) in surgical videos and proposes a framework called CurConMix.
🎯 What it does: This paper proposes a Circular Context Validation (CCV) framework that utilizes a self-check mechanism during the inference phase and query-specific prompts to enhance the performance of context-based learning in medical image segmentation.
🎯 What it does: A D-CAM method based on frequency domain decoupling is proposed, which utilizes weakly supervised image-level labels to train a classification network in the source domain to generate domain-invariant class activation maps (CAM), and then uses these pseudo-labels to train a segmentation model in the target domain, achieving generalizable medical image segmentation.
🎯 What it does: A dual-domain diffusion model D2Diff is proposed for synthesizing missing or defective contrast images from existing multi-contrast MRI.
🎯 What it does: A D-M model is proposed to embed spatial deformation in diffusion models for synthesizing enhanced T1 images of brain tumors from non-enhanced MRI.
Deep Association Multimodal Learning for Zero-shot Spatial Transcriptomics Prediction
Zhou, Yijing (East China Normal University), Wang, Yan (East China Normal University)
CodeGraph Neural NetworkTransformerLarge Language ModelPrompt EngineeringImageMultimodalityBiomedical Data
🎯 What it does: This paper proposes a deep associative multimodal framework that can predict gene expression of spatial transcriptomics from pathological images under zero-shot conditions.
Deep Knowledge-Infused Transformer for NSCLC Lymph Node Station Metastasis Prediction: Development of an AI-Powered Intraoperative Decision System
Zhao, Jie (Peking University), Dong, Bin (Peking University)
CodeGraph Neural NetworkTransformerBiomedical Data
🎯 What it does: This study investigates a Deep Knowledge Injection Transformer model (DKiT) for predicting lymph node station metastasis in patients with non-small cell lung cancer, and based on this, a real-time intraoperative decision support system was developed.
🎯 What it does: This study proposes DEFUSE-MS, a spatiotemporal heterogeneous graph neural network framework guided by deformation fields for the automatic detection of new T2 hyperintense lesions in multiple sclerosis.
🎯 What it does: A dual-guided multi-modal medical image registration framework DGMIR is proposed, achieving high-precision registration through modules such as multi-scale heterogeneous feature fusion, perspective feature reorganization, and modality removal.
🎯 What it does: A dynamic hierarchical graph Transformer framework called DHGFormer is proposed for generating and analyzing functional brain networks, thereby facilitating the diagnosis of brain diseases.
DiDGen: Diffusion-based Dual-task Synthesis for Dermoscopic Data Generation
Shentu, Junjie (Durham University), Al Moubayed, Noura (Durham University)
CodeClassificationSegmentationGenerationData SynthesisLarge Language ModelSupervised Fine-TuningDiffusion modelImageBiomedical Data
🎯 What it does: The DiDGen framework is proposed, utilizing diffusion models to simultaneously generate skin images and corresponding lesion masks under a single fine-tuning, addressing the issue of insufficient data.
CodeGenerationExplainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringImageTextBiomedical DataElectronic Health Records
🎯 What it does: Proposes the Diff-RRG framework, which utilizes longitudinal lesion-level difference maps to provide progression guidance for LLM, thereby generating radiology reports with more clinical semantics and interpretability.
🎯 What it does: This paper proposes DiffAtlas, which utilizes an image-mask diffusion model to jointly learn the joint distribution of images and segmentation masks during training, achieving Atlas segmentation based on generative AI.
🎯 What it does: This paper proposes DiffOSeg, a two-stage diffusion model that can simultaneously learn the consensus segmentation results of multiple experts and the individual preference segmentation results.
🎯 What it does: This study investigates a virtual staining method based on a conditional diffusion model that utilizes deep spectral clustering to generate high-quality multi-channel fluorescence images.
🎯 What it does: A multi-modal diffusion model (MMDM) is proposed to synthesize high-quality TOF-MRA images from multi-modal brain MRI (T1W, T2W, FLAIR).
Diffusion-Based User-Guided Data Augmentation for Coronary Stenosis Detection
Seo, Sumin (Medipixel, Inc.), Jung, Chung-Hwan (Medipixel, Inc.)
CodeClassificationObject DetectionData SynthesisConvolutional Neural NetworkDiffusion modelImageBiomedical Data
🎯 What it does: This paper proposes a user-guided data augmentation method based on diffusion models (DiGDA), which generates realistic narrow lesions in coronary angiography images by controlling vascular segmentation masks, original images, and text prompts, thereby achieving precise control over lesion severity (%DS). The synthetic images are used alongside real images to train a single-stage YOLO detection and severity classification model.
🎯 What it does: This paper proposes a virtual staining method based on a bridge diffusion model (RBDM), which maps Mueller matrix polarized images to H&E and fluorescence stained images, and constructs the first multimodal polarized pathology dataset.
DIGS: Dynamic CBCT Reconstruction using Deformation-Informed 4D Gaussian Splatting and a Low-Rank Free-Form Deformation Model
Huang, Yuliang (University College London), McClelland, Jamie R. (University College London)
CodeGaussian SplattingImageComputed Tomography
🎯 What it does: A 4D Gaussian splatting method based on deformation awareness (DIGS) is proposed for real-time reconstruction of dynamic CBCT images on each projection.
DINO Adapted to X-Ray (DAX): Foundation Models for Intraoperative X-Ray Imaging
Scheuplein, Joshua (Friedrich-Alexander-Universität), Kreher, Björn (Friedrich-Alexander-Universität)
CodeClassificationObject DetectionSegmentationConvolutional Neural NetworkTransformerContrastive LearningImageBiomedical Data
🎯 What it does: A large-scale intraoperative X-ray image dataset was constructed, and the DINO self-supervised learning framework was adapted as the DAX base model. Subsequently, evaluations were conducted on three downstream tasks: body region classification, metal implant segmentation, and screw detection.
🎯 What it does: This paper proposes a directionally adaptive shuffling-based Mamba model, DASMamba, specifically designed for medical image restoration. It achieves super-resolution, denoising, and synthesis of low-quality MRI, CT, and PET images through a U-shaped hierarchical structure combined with a state space model.
Disentangled and Interpretable Multimodal Attention Fusion for Cancer Survival Prediction
Eijpe, Aniek (Utrecht University), Silva, Wilson (Utrecht University)
CodeExplainability and InterpretabilityRepresentation LearningTransformerMultimodalityBiomedical Data
🎯 What it does: A multi-modal fusion framework based on attention, DIMAF, is proposed to integrate whole slide images (WSI) and transcriptomic data to predict cancer survival time, achieving representation decoupling through explicit separation of intra-modal and cross-modal interactions.
Distilling foundation models for robust and efficient models in digital pathology
Filiot, Alexandre (Owkin Inc), Olivier, Antoine (Owkin)
CodeCompressionComputational EfficiencyKnowledge DistillationTransformerContrastive LearningImageBiomedical Data
🎯 What it does: Using a teacher-student distillation approach, a digital pathology base model H-Optimus-0 with over one billion parameters was distilled into a ViT-Base model (H0-mini) with only 86M parameters, achieving significant compression in model size and inference cost.
DIY Challenge Blueprint: From Organization to Technical Realization in Biomedical Image Analysis
Klausmann, Leonard (OTH Regensburg), Palm, Christoph (OTH Regensburg)
CodeObject DetectionSegmentationPose EstimationImageBiomedical Data
🎯 What it does: This paper proposes and implements a DIY (self-built) challenge platform blueprint for biomedical image analysis, validated in the PhaKIR end-view challenge at MICCAI 2024.
🎯 What it does: By combining quantitative MRI parameters and physical forward models, two physics-constrained synthetic data methods (qATLAS and qSynth) are proposed and used to train stroke lesion segmentation models to enhance cross-domain generalization capabilities.
🎯 What it does: A dual-prompt driven network DpDNet was designed and implemented for the unified segmentation of multiple cancer types in whole-body tumors from PET-CT.
🎯 What it does: The DSFC framework is proposed, which adapts the basic model of medical imaging to the domain through global and local deformation perturbations and a self-feedback loop, and introduces hard sample adaptive loss to enhance training stability.
🎯 What it does: A dual correlation self-attention Mamba model is proposed, utilizing neighboring frame correlation and diastolic frame correlation to identify microvascular occlusion in non-contrast cardiac magnetic resonance imaging.
Dual Knowledge-Aware Guidance for Source-Free Domain Adaptive Fundus Image Segmentation
Chen, Yu (Shanghai Key Laboratory of Trustworthy Computing, East China Normal University), Cao, Guitao (Ant Group)
CodeSegmentationDomain AdaptationImage
🎯 What it does: This paper proposes a Dual Knowledge-Guided (DKG) method for source-independent domain adaptation of retinal image segmentation, enhancing segmentation performance in the target domain.
🎯 What it does: This paper proposes the DDCWISS network, which utilizes weakly supervised image-level labels to achieve class-incremental segmentation of pathological regions, and reduces the cost of pixel-level annotations and catastrophic forgetting through dual-branch dynamic coupling and dual-path supervision.
🎯 What it does: A Dual-Stream Multi-Frequency Fusion Network (DSMFN) is proposed for dynamic functional connectivity analysis to achieve classification of brain disorders (MCI, ASD).
DualPrompt-MedCap: A Dual-Prompt Enhanced Approach for Medical Image Captioning
Zhao, Yining (University of Technology Sydney), Braytee, Ali (University of Technology Sydney)
CodeRecognitionGenerationConvolutional Neural NetworkSupervised Fine-TuningPrompt EngineeringVision Language ModelImageTextMultimodalityMagnetic Resonance Imaging
🎯 What it does: This paper proposes DualPrompt-MedCap, a dual-prompt medical image captioning framework that combines modality awareness and question guidance.
🎯 What it does: The DuoDent framework is proposed, utilizing a dual-stream diffusion model to generate high-precision dental point clouds, and reconstructing smooth dental meshes through regularized normal consistency and directional consistency.
🎯 What it does: A self-supervised optimization framework Dyna3DGR is proposed, which combines explicit 3D Gaussian representation with implicit neural motion fields to estimate 4D cardiac motion in real-time;
🎯 What it does: This paper proposes a LN segmentation baseline model using Dynamic Gradient Sparsification Training (DGST) under few-shot conditions, enhancing the segmentation performance of CT neck and mediastinal LN through pre-training and fine-tuning.
🎯 What it does: A Dynamic Aware Implicit Neural Representation (DA-INR) framework is proposed for unsupervised dynamic MRI reconstruction, achieving fast convergence and high-quality reconstruction by combining hash coding and normalized space modeling.
🎯 What it does: A probabilistic student-teacher model called EchoingECG has been developed to predict echocardiographic (ECHO) related metrics using ECG signals.
EchoViewCLIP: Advancing Video Quality Control through High-performance View Recognition of Echocardiography
Song, Shanshan (Hong Kong University of Science and Technology), Li, Xiaomeng (Southern Medical University)
CodeRecognitionAnomaly DetectionVision Language ModelContrastive LearningVideoMultimodalityUltrasound
🎯 What it does: This paper proposes the EchoViewCLIP framework, achieving fine recognition of 38 categories of cardiac ultrasound views, and integrates OOD detection and quality assessment.
Edge-Aware Token Halting for Efficient and Accurate Medical Image Segmentation
Guo, Yuhao (Institute of Intelligent Machines, Chinese Academy of Sciences), Cheng, Erkang (Institute of Intelligent Machines, Chinese Academy of Sciences)
🎯 What it does: This paper proposes an efficient segmentation framework HRViT based on Vision Transformer, which achieves high-precision semantic segmentation of medical images through dynamic edge token stopping and reconstruction.
🎯 What it does: This paper proposes a handwriting-supervised medical image segmentation method called EFFDNet, which enhances foreground recognition capability by utilizing the foreground-background semantics in handwriting.
🎯 What it does: The DeCoCT method is proposed, which decomposes the causal relationship between clinical terms and answers in medical visual question answering (Med-VQA), uses adversarial contrastive training to eliminate language bias, and introduces a Key Region Capture Module (KRCM) to enhance the model's focus on visual key information.
🎯 What it does: An end-to-end two-stage 3D tooth landmark detection framework, ToothLDNet, is proposed, which can directly locate teeth and detect landmarks from 3D mandibular models without the need for tooth segmentation labels.
🎯 What it does: In the dynamic endoscopic scene, we propose Endo-4DGX, which combines illumination embedding, region-aware enhancement, and spatial-aware correction within a Gaussian expansion framework to achieve 3D reconstruction and illumination correction under uneven lighting.
🎯 What it does: Developed Endo-FASt3r, a self-supervised learning-based monocular depth and pose estimation framework, utilizing two foundational models (Depth Anything V2 and Reloc3rX) for end-to-end estimation.
Endo-GSMT: Endoscopic Monocular Scene Reconstruction with Dynamic Gaussian Splatting and Motion Tracking
Gou, Hao (Hangzhou Dianzi University), Luo, Huoling (Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences)
CodeGaussian SplattingSimultaneous Localization and MappingVideo
🎯 What it does: A dynamic surgical scene reconstruction method based on monocular endoscopic video is proposed, utilizing 3D Gaussian Splatting to achieve high-quality 3D reconstruction.
🎯 What it does: Perform parameter-efficient fine-tuning of Video Depth Anything on endoscopic videos to achieve spatial accuracy and temporal consistency in depth estimation.
EndoFlow-SLAM: Real-Time Endoscopic SLAM with Flow-Constrained Gaussian Splatting
Wu, Taoyu (Xi'an Jiaotong Liverpool University), Li, Haoang (Hong Kong University of Science and Technology (Guangzhou))
CodePose EstimationDepth EstimationOptimizationGaussian SplattingSimultaneous Localization and MappingOptical FlowBiomedical Data
🎯 What it does: A real-time endoscopic SLAM system (EndoFlow-SLAM) based on 3D Gaussian Splatting with optical flow constraints is proposed, achieving accurate camera pose estimation and high-quality dense reconstruction.
🎯 What it does: EndoGen is proposed, a conditional autoregressive endoscopic video generation framework that can generate high-quality, temporally consistent endoscopic videos based on pathological categories.
🎯 What it does: A two-stage framework for removing artifacts from endoscopic images is proposed, which first suppresses specular reflections and then fills in the diffuse reflection areas to enhance segmentation performance.
🎯 What it does: A dual-stream cascade encoding-decoding network is proposed to achieve multi-focus image fusion for endoscopy, thereby extending the depth of field.
🎯 What it does: The first publicly available high-resolution lumbar endplate 3D QCT dataset, Endplate3D-QCT, has been proposed, along with pixel-level annotations and an evaluation framework.
Enforcing Geometric Constraints of Surface Normal and Pose for Self-supervised Monocular Depth Estimation on Laparoscopic Images
Li, Wenda (Nagoya University), Mori, Kensaku (Aichi Institute of Technology)
CodePose EstimationDepth EstimationVideo
🎯 What it does: This paper proposes a self-supervised monocular depth estimation method that enhances depth estimation quality on laparoscopic images by introducing surface normal estimation, distance uncertainty, and a 4D score volume to achieve geometric constraints.
🎯 What it does: An adaptive uncertainty estimation network based on video and audio multimodal data, AUSTIN, is proposed to enhance the accuracy and reliability of stroke triage in the emergency room.
🎯 What it does: A dataset of expert annotated reports for four imaging modalities (cardiac MRI, abdominal ultrasound, head CT, and CT pulmonary angiography) was constructed, and based on this, RadGraph was fine-tuned for modality-specific tasks;
🎯 What it does: Utilizing a multi-instance learning framework combined with supervised contrastive learning and propensity score matching to eliminate patient-specific biases in WSI, thereby improving the classification performance of soft tissue sarcoma subtypes.
Enhancing WSI-Based Survival Analysis with Report-Auxiliary Self-Distillation
Wang, Zheng (Xiamen University), Wang, Liansheng (Southern Medical University)
CodeKnowledge DistillationTransformerLarge Language ModelSupervised Fine-TuningImageTextBiomedical Data
🎯 What it does: This paper proposes a WSI survival analysis framework called Rasa, based on report-assisted self-distillation, which extracts precise report text using LLM to guide feature selection and data augmentation, thereby improving the survival prediction performance of WSI.
Lemke, Nick (Technical University of Darmstadt), Mukhopadhyay, Anirban (Carl Zeiss AG)
CodeSegmentationFederated LearningSafty and PrivacyComputational EfficiencyImageBiomedical DataUltrasound
🎯 What it does: This paper proposes FedNCA—a federated learning framework based on a lightweight neural cellular automaton (Med-NCA) that can perform medical image segmentation tasks on edge devices with low computing power and low bandwidth (such as mobile phones), while supporting homomorphic encryption to ensure secure aggregation on the server.
ESPNet: Edge-Aware Feature Shrinkage Pyramid for Polyp Segmentation
Toman, Raneem (University of Leeds), Ali, Sharib (University of Leeds)
CodeSegmentationTransformerImage
🎯 What it does: This paper presents ESPNet, a Transformer-based edge-aware feature shrinkage pyramid network for polyp segmentation in multi-center and diverse populations.
🎯 What it does: We propose EUReg, an end-to-end real-time 2D-3D ultrasound registration framework that addresses the shortcomings of existing methods in terms of accuracy, efficiency, and overfitting.
Explainable Classifier for Malignant Lymphoma Subtyping via Cell Graph and Image Fusion
Nishiyama, Daiki (Institute of Science Tokyo), Sakuma, Jun (Institute of Science Tokyo)
CodeClassificationExplainability and InterpretabilityConvolutional Neural NetworkGraph Neural NetworkMixture of ExpertsImageMultimodalityBiomedical Data
🎯 What it does: An interpretable multimodal MIL framework is proposed, utilizing cell graphs and image fusion to classify three subtypes of malignant lymphoma (DLBCL, FL, Reactive), and provides class-level ROI and explanations of cell frequency and spatial distribution.
Explainable Integrative Bipartite Graph Convolutional Neural Network for Predicting Ejection Fraction in Echocardiography
Lee, Seungeun (Klleon), Kang, Mingon (University of Nevada Las Vegas)
CodeExplainability and InterpretabilityConvolutional Neural NetworkGraph Neural NetworkAuto EncoderVideoMultimodalityTabularUltrasound
🎯 What it does: An interpretable bilateral graph convolutional neural network, IBi-GNN, is proposed to integrate cardiac ultrasound videos and demographic features (age, gender, BMI) to predict left ventricular ejection fraction.
CodeSegmentationConvolutional Neural NetworkMixture of ExpertsImageTextMultimodalityBiomedical DataComputed Tomography
🎯 What it does: This paper studies a text-enhanced mixture of experts model, TextMoE, for semi-supervised medical image segmentation of combined CT and X-ray images.
Exploring the Design Space of 3D MLLMs for CT Report Generation
Baharoon, Mohammed (Vector Institute for Artificial Intelligence), Wang, Bo (Vector Institute for Artificial Intelligence)
CodeGenerationTransformerLarge Language ModelSupervised Fine-TuningMultimodalityBiomedical DataComputed Tomography
🎯 What it does: This paper systematically evaluates the design space of 3D multimodal large language models (MLLM) in CT report generation and proposes two knowledge enhancement methods to improve the completeness and quality of reports.
Exposing and Mitigating Calibration Biases and Demographic Unfairness in MLLM Few-Shot In-Context Learning for Medical Image Classification
Shen, Xing (McGill University), Arbel, Tal (McGill University)
CodeClassificationTransformerLarge Language ModelImageMultimodalityBiomedical DataMagnetic Resonance Imaging
🎯 What it does: This study investigates the calibration bias and fairness issues of multimodal large language models in few-shot contextual learning for medical imaging, and proposes a training-free second-order calibration method called CALIN.
F2PASeg: Feature Fusion for Pituitary Anatomy Segmentation in Endoscopic Surgery
Chen, Lumin (Centre for Artificial Intelligence and Robotics, Hong Kong Institute of Science Innovation, Chinese Academy of Sciences), Liu, Hongbin (Centre for Artificial Intelligence and Robotics, Hong Kong Institute of Science Innovation, Chinese Academy of Sciences)
CodeSegmentationTransformerImageBiomedical Data
🎯 What it does: Proposes F2PASeg, which combines a Feature Fusion module for real-time anatomical structure semantic segmentation in endoscopic pituitary surgery.
🎯 What it does: The CAMOS framework is proposed, which uses a conditional autoregressive multiscale model to directly predict the optimal 3D appearance post-surgery from the preoperative facial deformity of patients.
CodeTransformerMixture of ExpertsVision Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: The Fair-MoE framework is proposed, which enhances fairness and diagnostic effectiveness in medical vision-language models through two modules: FO-MoE and FOL.
🎯 What it does: The GlaucoDiff model is proposed, which can achieve bidirectional generation of healthy and glaucoma retinal images by controlling the vertical cup-to-disc ratio (vCDR), thereby enriching the dataset and enhancing diagnostic fairness.
🎯 What it does: A self-supervised multi-view super-resolution network called tripleSR is proposed, which fuses two orthogonal low-resolution MRI images using sparse coordinate loss to generate a unified high-resolution image.
🎯 What it does: A self-supervised medical image segmentation framework FDAS based on foundational model distillation and anatomy structure-aware multi-task learning is proposed.
🎯 What it does: This paper proposes an interpretable multi-sequence MRI enhancement framework FDF-VQVAE, which achieves image denoising, super-resolution, and motion artifact removal through frequency domain feature separation and fusion.
🎯 What it does: This study investigates brain tumor segmentation in a federated learning environment and proposes the FedAMM framework to address the issue of arbitrary missing modalities.