🎯 What it does: Designed and implemented the DB-SAM dual-branch (ViT + convolution) framework, transforming SAM to achieve unified high-quality medical image segmentation.
Debiased Noise Editing on Foundation Models for Fair Medical Image Classification
Jin, Ruinan (University of British Columbia), Li, Xiaoxiao (University Of British Columbia)
CodeClassificationTransformerSupervised Fine-TuningImageBiomedical Data
🎯 What it does: This paper proposes a method to eliminate false associations related to sensitive attributes by adding learnable denoising noise (DNE) to medical images, and implements fair embedding in a black-box base model API.
🎯 What it does: The paper proposes a method to improve semi-supervised medical image segmentation through decoupled training and worst-case aware learning.
🎯 What it does: Reconstruct the three-dimensional volume of transparent samples such as cells using multi-focal microscopy images through the Deep Image Prior (DIP) framework.
Deform3DGS: Flexible Deformation for Fast Surgical Scene Reconstruction with Gaussian Splatting
Yang, Shuojue (National University of Singapore), Jin, Yueming (University of Oxford)
CodeGaussian SplattingVideoPoint Cloud
🎯 What it does: The Deform3DGS framework is proposed, achieving rapid deformable tissue reconstruction by using 3D Gaussian Splatting and a flexible deformation model in surgical videos.
🎯 What it does: This paper proposes a deep perception-based endoscopic video restoration framework called DAEVI, which achieves more realistic 3D spatial detail recovery by directly inferring depth from visual features, dual-modal paired channel fusion, and a depth-enhanced discriminator.
🎯 What it does: This paper proposes a deep-driven geometric prompt learning network for laparoscopic liver landmark detection, named D2GPLand, and releases a new L3D dataset.
DermaVQA: A Multilingual Visual Question Answering Dataset for Dermatology
Yim, Wen-wai (Microsoft Health AI), Xia, Fei (University of Washington)
CodeTransformerLarge Language ModelVision Language ModelImageTextMultimodality
🎯 What it does: A multilingual (Chinese and English) dermatology visual question answering (Dermatology VQA) dataset called DermaVQA has been constructed, and baseline evaluations of three state-of-the-art multimodal large language models (Gemini-Pro Vision, GPT-4-vision-preview, LLaVA-FT+GPT-4) have been conducted on this dataset to explore the task of generating doctor responses.
DES-SAM: Distillation-Enhanced Semantic SAM for Cervical Nuclear Segmentation with Box Annotation
Huang, Lina (Central South University), Liu, Jianfeng (Central South University)
CodeSegmentationKnowledge DistillationImage
🎯 What it does: A box supervision cervical nucleus segmentation network DES-SAM based on self-distillation prompts is proposed, aiming to reduce manual annotation costs and improve the model's generalization ability.
🎯 What it does: This paper studies a model to enhance the generalization of medical image segmentation under single-source domain transfer—Decoupled Segment Anything Model (DeSAM), which achieves automated segmentation by decoupling the association between prompts and masks.
Design as Desired: Utilizing Visual Question Answering for Multimodal Pre-training
Su, Tongkun (Shenzhen Institute of Advanced Technology Chinese Academy of Sciences), Hu, Ying (Shenzhen Institute of Advanced Technology Chinese Academy of Sciences)
CodeClassificationSegmentationGenerationTransformerVision Language ModelContrastive LearningImageTextMultimodalityUltrasound
🎯 What it does: This paper proposes the use of Visual Question Answering (VQA) for pre-training on medical multimodal data and designs multi-granularity question-answer pairs to guide the model in learning pathological features.
🎯 What it does: A multi-task video polyp segmentation framework based on diffusion models, Diff-VPS, is proposed, which enhances the segmentation capability for high-fidelity colored polyps through a temporal reasoning module.
🎯 What it does: This paper proposes the Diff3Dformer model, which utilizes a diffusion autoencoder to extract semantic features from CT slices. It then generates slice prototypes through spherical K-means clustering and uses the cluster numbers as auxiliary input to a clustering attention Transformer to complete 3D CT classification.
🎯 What it does: This paper trains a score diffusion model solely on non-motion CT images to compute the exact likelihood of any CT image, using the likelihood gradient for gradient descent optimization of motion parameters, thereby achieving CT motion compensation.
🎯 What it does: The study proposes a Microsoft morphological filter in medical image segmentation and integrates it into the network's loss function or final layer.
CodeGenerationExplainability and InterpretabilityKnowledge DistillationDiffusion modelAuto EncoderImageBiomedical DataComputed Tomography
🎯 What it does: This paper proposes DiffExplainer, a counterfactual generation framework that combines diffusion autoencoders and teacher-student learning to reveal the decision basis of black-box models in medical CT image classification.
🎯 What it does: This paper proposes a semi-supervised medical image segmentation framework named DiffRect, which combines pseudo-label correction with a latent diffusion model to enhance segmentation performance.
🎯 What it does: This paper proposes DiffuseReg, which achieves unsupervised deformable image registration by gradually denoising the displacement field using a diffusion model.
Diffusion as Sound Propagation: Physics-inspired Model for Ultrasound Image Generation
Domínguez, Marina (Technical University of Munich), Azampour, Mohammad Farid (Technical University of Munich)
CodeGenerationData SynthesisDiffusion modelImageBiomedical DataUltrasoundPhysics Related
🎯 What it does: The study introduces a physics-inspired B-Maps noise scheduler in ultrasound image generation to improve diffusion models, better simulating sound wave attenuation and achieving more realistic synthetic B-mode ultrasound images.
🎯 What it does: To address the limitations of body composition analysis caused by field of view (FOV) truncation in chest CT scans, this paper proposes an image outpainting method based on diffusion models to automatically restore truncated slices, thereby enabling complete measurements of muscle and fat areas.
Diffusion-Enhanced Transformation Consistency Learning for Retinal Image Segmentation
Li, Xiang (Nanyang Technological University), Duan, Lixin (University of Electronic Science and Technology of China)
CodeSegmentationDiffusion modelImage
🎯 What it does: A semi-supervised retinal image segmentation framework named DiffTCL is proposed, which enhances segmentation accuracy under limited annotations through self-supervised diffusion pre-training and transformation consistency learning.
🎯 What it does: Using CBCT/3dMD images, 328 3D facial soft tissue landmark points were extracted through the MediaPipe 2D landmark model, and then these landmark points were diagnosed using the Diagnosis-Reconstruction Transformer (DiRecT), supplemented by a reconstruction task and teacher-student semi-supervised learning to enhance model performance.
CodeClassificationExplainability and InterpretabilityGraph Neural NetworkGraphBiomedical DataAlzheimer's Disease
🎯 What it does: The Disentangled Attention Graph Neural Network (DAGNN) model is proposed for the diagnosis of Alzheimer's Disease (ADD) and Mild Cognitive Impairment (MCI) based on structural networks.
Distributionally-Adaptive Variational Meta Learning for Brain Graph Classification
Du, Jing (Macquarie University), Giral, Alexis (Systemethix)
CodeClassificationMeta LearningGraph Neural NetworkAuto EncoderGraphBiomedical Data
🎯 What it does: A distribution-adaptive variational meta-learning framework (DAML) is proposed for brain image classification, combining graph representation learning and distribution-adaptive variational meta-learning to address data drift and small sample issues.
🎯 What it does: A method named DnFPlane is proposed to achieve efficient and high-definition 4D reconstruction of deformable tissues, particularly addressing the issues of depth distortion and dynamic blur caused by occlusion from surgical instruments.
🎯 What it does: This paper proposes a domain-adaptive unsupervised cancer detection method (DAUD) that utilizes autoencoders and latent variables to identify malignant and unknown cases in skin tissue slices.
🎯 What it does: Developed the RL4Seg framework, utilizing reinforcement learning to achieve unsupervised 2D ultrasound cardiac image segmentation and uncertainty estimation, significantly reducing the need for expert annotations.
Double-tier Attention based Multi-label Learning Network for Predicting Biomarkers from Whole Slide Images of Breast Cancer
Wang, Mingkang (Dalian University of Technology), Xu, Hongming (Dalian University Of Technology)
CodeClassificationConvolutional Neural NetworkTransformerImageBiomedical Data
🎯 What it does: This paper proposes a Dual-layer Attention Multi-label Learning Network (DAMLN) that can simultaneously predict the expression status of estrogen receptor (ER), progesterone receptor (PR), and human epidermal growth factor receptor 2 (HER2) using only H&E stained whole slide images (WSI) of breast cancer.
DRIM: Learning Disentangled Representations from Incomplete Multimodal Healthcare Data
Robinet, Lucas (IUCT-Oncopole-Institut Claudius Regaud), Cohen-Jonathan Moyal, Elizabeth (IUCT-Oncopole-Institut Claudius Regaud)
CodeOptimizationRepresentation LearningTransformerContrastive LearningMultimodalityBiomedical DataMagnetic Resonance ImagingElectronic Health Records
🎯 What it does: Proposes the DRIM framework, which decomposes incomplete multimodal medical data using shared and unique encoders, and further integrates them to achieve survival prediction.
CodeClassificationConvolutional Neural NetworkMultimodalityBiomedical Data
🎯 What it does: A multi-modal fusion network based on dynamic selection and clinical enhancement (DSCENet) is proposed for the precise classification of subtypes of myeloproliferative neoplasms (MPN).
🎯 What it does: Pseudo-labels are generated using point annotations and dynamically optimized, eliminating the need for pixel-level annotations in nuclear segmentation.
Dynamic Single-Pixel Imaging on an Extended Field of View without Warping the Patterns
Maitre, Thomas (Universite Claude Bernard Lyon 1), Sdika, Michaël (Universite Claude Bernard Lyon 1)
CodeRestorationOptical FlowImage
🎯 What it does: This paper addresses the artifact problem caused by motion in dynamic single-pixel imaging (SPI) and proposes a complete decoupling framework that does not convolve distort the light patterns in an extended field of view (FOV);
EchoMEN: Combating Data Imbalance in Ejection Fraction Regression via Multi-Expert Network
Lai, Song (City University of Hong Kong), Meng, Gaofeng (Chinese Academy of Sciences)
CodeData-Centric LearningMixture of ExpertsContrastive LearningVideoBiomedical DataUltrasound
🎯 What it does: This paper proposes EchoMEN, a multi-expert network designed to address the data imbalance problem in the regression of ventricular ejection fraction (EF).
EchoNarrator: Generating natural text explanations for ejection fraction predictions
Thomas, Sarina (University of Oslo), Ben-Yosef, Guy (GE Healthcare)
CodeGenerationExplainability and InterpretabilityConvolutional Neural NetworkGraph Neural NetworkLarge Language ModelSupervised Fine-TuningVideoTextUltrasoundChain-of-Thought
🎯 What it does: An end-to-end system was constructed to detect the left ventricle contour using multi-frame GCN from ultrasound videos and directly predict the ejection fraction (EF), followed by generating clinically readable natural language explanations through geometric attributes.
EchoNet-Synthetic: Privacy-preserving Video Generation for Safe Medical Data Sharing
Reynaud, Hadrien (Imperial College London), Kainz, Bernhard (Imperial College London)
CodeGenerationData SynthesisSafty and PrivacyDiffusion modelAuto EncoderVideoBiomedical DataUltrasound
🎯 What it does: A complete end-to-end process has been designed and implemented to generate high-fidelity, long-duration synthetic cardiac ultrasound videos using a latent video diffusion model, with privacy filtering to ensure data de-identification.
🎯 What it does: A learning-based two-stage coarse-fine point tracking model called EchoTracker is proposed, specifically designed for tracking the trajectories of myocardial points in cardiac ultrasound, and utilizes these trajectories to calculate global longitudinal strain (GLS).
EgoSurgery-Phase: A Dataset of Surgical Phase Recognition from Egocentric Open Surgery Videos
Fujii, Ryo (Keio University), Kajita, Hiroki (Keio University School of Medicine)
CodeRecognitionTransformerAuto EncoderVideo
🎯 What it does: This paper constructs the first publicly available subjective perspective open surgery video dataset Egosurgery-Phase and proposes the gaze-guided mask autoencoder GGMAE for surgical phase recognition.
🎯 What it does: This paper proposes a 3D medical image segmentation framework called EM-Net based on Mamba, which integrates Channel Squeeze-Excitation Blocks (CSRM), Frequency Domain Learning Layers (EFL), and a Mamba decoder to achieve efficient segmentation.
🎯 What it does: An online learning framework is proposed, utilizing three mechanisms: linear, dynamic, and selective memory to efficiently process continuous streams of medical data, alleviate catastrophic forgetting, and significantly improve multi-organ and tumor segmentation performance.
Embryo Graphs: Predicting Human Embryo Viability from 3D Morphology
He, Chloe (University College London), Vasconcelos, Francisco (Imperial College London)
CodeClassificationSegmentationConvolutional Neural NetworkGraph Neural NetworkImageMeshGraphBiomedical Data
🎯 What it does: A non-invasive 3D reconstruction and graphical representation process is proposed, utilizing the generated embryonic graph structure to train a graph neural network to predict the live birth rate of embryos.
Enable the Right to be Forgotten with Federated Client Unlearning in Medical Imaging
Deng, Zhipeng (Hong Kong University of Science and Technology), Chen, Hao (Harvard Medical School)
CodeFederated LearningSafty and PrivacyContrastive LearningImageBiomedical Data
🎯 What it does: Proposed the Federated Client Unlearning (FCU) framework, which implements the 'right to be forgotten' for clients in medical image federated learning;
Enabling Text-free Inference in Language-guided Segmentation of Chest X-rays via Self-guidance
Ye, Shuchang (University of Sydney), Kim, Jinman (University of Sydney)
CodeObject DetectionSegmentationConvolutional Neural NetworkLarge Language ModelImageMultimodality
🎯 What it does: A self-guided segmentation framework SGSeg is proposed, which uses language guidance only during the training phase, allowing for infection area segmentation of chest X-ray images without text during inference.
EndoDAC: Efficient Adapting Foundation Model for Self-Supervised Depth Estimation from Any Endoscopic Camera
Cui, Beilei (Chinese University of Hong Kong), Ren, Hongliang (Chinese University of Hong Kong)
CodeDepth EstimationTransformerOptical FlowVideo
🎯 What it does: Proposed the EndoDAC framework, which efficiently adapts deep foundational models based on Vision Transformer for monocular depth estimation of any endoscopic camera in a self-supervised manner.
🎯 What it does: A self-supervised monocular 3D scene reconstruction framework based on NeRF, called EndoSelf, is proposed, which completes the 3D reconstruction of deformable tissues using endoscopic video without depth supervision.
🎯 What it does: This paper proposes an energy model-controlled radiology report generation method (ECRG), which incorporates pre-trained expert systems or medical knowledge during the inference phase to impose energy constraints on the report generation process, enabling controllable generation.
🎯 What it does: An end-to-end image-based method is proposed, utilizing the geometric model of surgical instruments to achieve scale-aware depth estimation from monocular endoscopic images.
Enhancing Gene Expression Prediction from Histology Images with Spatial Transcriptomics Completion
Mejia, Gabriel (Universidad de los Andes), Arbeláez, Pablo (Universidad de los Andes)
CodeGraph Neural NetworkTransformerImageBiomedical Data
🎯 What it does: This paper constructs 26 standardized Visium Spatial Transcriptomics datasets (SpaRED) and proposes a Transformer-based reference-independent missing value imputation model, SpaCKLE, to enhance gene expression prediction performance.
Enhancing Human-Computer Interaction in Chest X-ray Analysis using Vision and Language Model with Eye Gaze Patterns
Kim, Yunsoo (University College London), Wu, Honghan (University College London)
CodeTransformerVision Language ModelImageMultimodalityBiomedical DataComputed Tomography
🎯 What it does: This paper inputs the eye-tracking heatmaps of radiologists along with chest X-ray images into a visual-language model to enhance the accuracy of multi-task diagnosis.
🎯 What it does: A medical image segmentation framework called TextDiff based on diffusion models is proposed, which utilizes diagnostic text annotations to enhance visual semantic representation, thereby achieving efficient segmentation with few labeled samples.
🎯 What it does: A two-stage sampling pipeline based on foundational model features (DataDIVA) is proposed, which enhances the model's generalization performance on both internal and external datasets through diversity and balanced sampling on retinal images.
🎯 What it does: By using self-supervised mask pre-training and synthetic lesion augmentation, the accuracy of new lesion segmentation in multiple sclerosis is improved; training is conducted on a VNet-based segmentation model using both real and synthetic lesions as input; model parameters are pre-trained through self-supervised tasks, and consistency loss and boundary loss are added in the segmentation task; multi-fold cross-validation evaluation is implemented.
🎯 What it does: Proposed and implemented BrLP, a 3D brain MRI spatiotemporal disease progression prediction framework utilizing latent diffusion models, ControlNet, auxiliary models, and LAS technology.
Ensemble of Prior-guided Expert Graph Models for Survival Prediction in Digital Pathology
Ramanathan, Vishwesh (University of Toronto), Martel, Anne L. (University of Toronto)
CodeClassificationSegmentationGraph Neural NetworkImageBiomedical Data
🎯 What it does: A directed edge attribute tissue subgraph is constructed using tissue and TIL segmentation priors, and a specialized graph attention network is trained on each subgraph. Finally, the expert models are linearly integrated to obtain the final survival prediction.
🎯 What it does: This paper presents DISYRE v2, an unsupervised anomaly detection framework that combines cold diffusion recovery and synthetic anomaly generation for abnormal localization in brain MRI.
🎯 What it does: A lightweight end-to-end cardiac ultrasound image frame-volume registration network CU-Reg has been designed and implemented to register real-time two-dimensional ultrasound images with preoperative three-dimensional ultrasound volumes, providing a complete surgical navigation view.
Epileptic Seizure Detection in SEEG Signals using a Unified Multi-scale Temporal-Spatial-Spectral Transformer Model
Li, Zhuoyi (Northwestern Polytechnical University), Zhang, Tuo (Northwestern Polytechnic University)
CodeClassificationAnomaly DetectionTransformerTime SeriesBiomedical Data
🎯 What it does: A unified multi-scale spatio-temporal frequency domain Transformer framework (CE-TSS-Transformer) is proposed for detecting long-term SEEG signals of epilepsy seizures, and a novel LTSZ epilepsy seizure SEEG dataset is constructed.
🎯 What it does: An unsupervised image harmonization framework called ESPA is proposed, which generates matched data by simulating scanner effects and using structure-preserving enhancement methods to achieve debiasing of multi-scanner images.
🎯 What it does: This paper proposes an implicit neural representation method based on grid hash encoding (HashEnc) for continuously estimating the orientation distribution function (ODF) field in high-resolution diffusion magnetic resonance scans.
🎯 What it does: Evaluate the quality of brain MRI generators and propose a unified framework for preprocessing, implementation, and anatomical plausibility assessment.
CodeClassificationExplainability and InterpretabilityVision Language ModelImageBiomedical Data
🎯 What it does: A concept bottleneck model based on evidence reasoning (evi-CEM) is proposed for skin disease diagnosis; simultaneously, a concept correction mechanism (using CAV to correct concept biases generated by VLM) and an uncertainty-based intervention strategy are designed.
Explainable vertebral fracture analysis with uncertainty estimation using differentiable rule-based classification
Wåhlstrand Skärström, Victor (Chalmers University of Technology), Häggström, Ida (Chalmers University of Technology)
CodeClassificationExplainability and InterpretabilityConvolutional Neural NetworkFlow-based ModelImageBiomedical Data
🎯 What it does: An end-to-end interpretable spinal compression fracture assessment system (XVFA) is proposed, which first locates the vertebrae and predicts the positions of key points, and then uses a differentiable Genant semi-quantitative rule for fracture level and morphology classification, providing uncertainty estimates.
Exploring Spatio-Temporal Interpretable Dynamic Brain Function with Transformer for Brain Disorder Diagnosis
Li, Lanting (Northeastern University), Zaiane, Osmar R. (Amii)
CodeClassificationExplainability and InterpretabilityTransformerTime SeriesBiomedical DataMagnetic Resonance Imaging
🎯 What it does: An end-to-end Transformer framework called BISTformer is proposed for spatiotemporal clustering of brain functional modules and diagnosing brain diseases.
🎯 What it does: Proposes F2TNet, which infers using low-cost T1w MRI by transferring fMRI knowledge to T1w MRI and simultaneously predicting multiple phenotypes.
🎯 What it does: Proposes PatchAlign, which aligns skin image patches with clinical text labels through Masked Graph Optimal Transport to enhance the accuracy and fairness of skin disease classification.
CodeSegmentationData SynthesisDiffusion modelImagePoint CloudBiomedical Data
🎯 What it does: Generate synthetic SLO retinal images that conform to boundary constraints through Point-Image Diffusion, and combine the synthetic data with real data in an Equal-Scale manner for training a fair segmentation model.
🎯 What it does: By quantizing the weights of deep models based on power binary, using fairness scores derived from Hessian, selectively quantizing certain weights to enhance diagnostic fairness among different populations while maintaining overall accuracy.
FALFormer: Feature-aware Landmarks self-attention for Whole-slide Image Classification
Bui, Doanh C. (Korea University), Kwak, Jin Tae (Korea University)
CodeClassificationTransformerImage
🎯 What it does: A Transformer-based FALFormer model is proposed, utilizing Feature-Aware Landmarks Self-Attention (FALSA) for efficient classification on complete WSI.
FastSAM3D: An Efficient Segment Anything Model for 3D Volumetric Medical Images
Shen, Yiqing (Johns Hopkins University), Unberath, Mathias (University of Arkansas)
CodeSegmentationComputational EfficiencyKnowledge DistillationTransformerVision Language ModelBiomedical DataMagnetic Resonance ImagingComputed Tomography
🎯 What it does: Introducing FastSAM3D, an efficient Segment Anything model specifically designed for 3D medical imaging, supporting fast interactive voxel-level segmentation;
🎯 What it does: This study evaluates and compares the consistency of feature extractors trained on ImageNet and RadImageNet in the assessment of medical image generation models (StyleGAN2), and verifies the effectiveness of augmentation techniques such as DiffAugment.
Feature Fusion Based on Mutual-Cross-Attention Mechanism for EEG Emotion Recognition
Zhao, Yimin (Southwest Jiaotong University), Gu, Jin (Southwest Jiaotong University)
CodeClassificationRecognitionConvolutional Neural NetworkTime SeriesBiomedical Data
🎯 What it does: This paper proposes a feature fusion method based on the Mutual Cross Attention mechanism (MCA), combined with a custom 3D-CNN, to classify emotions using DE and PSD features from EEG.
🎯 What it does: A training-free GBM segmentation framework called GBMSeg is proposed, which can achieve automatic segmentation of the glomerular basement membrane using a single annotated reference TEM image.
🎯 What it does: This paper proposes FedEvi, an evidence-based federated learning method that dynamically adjusts aggregation weights using global generalization gaps and local reliability to enhance the generalization performance of multi-center medical image segmentation.
CodeSegmentationFederated LearningSafty and PrivacyTransformerImageBiomedical DataMagnetic Resonance Imaging
🎯 What it does: Proposes the FedFMS framework, introducing SAM and its efficient variant MSA into federated learning for multi-center medical image segmentation, and conducts experimental validation on four types of non-IID datasets.
🎯 What it does: A federated learning framework named FedIA is proposed to address the issue of uneven labeling completeness among different clients in medical image segmentation.
🎯 What it does: Proposes the FedMedICL framework and benchmark for simultaneously evaluating label, demographic, and temporal distribution shifts in federated medical imaging, and simulates disease spread through continual learning.
FedMLP: Federated Multi-Label Medical Image Classification under Task Heterogeneity
Sun, Zhaobin (Huazhong University of Science and Technology), Yan, Zengqiang (Huazhong University of Science and Technology)
CodeClassificationFederated LearningImageBiomedical Data
🎯 What it does: The study proposes the FedMLP model to address task heterogeneity in multi-label medical image classification, solving the problem of missing labels.
Few-shot Adaptation of Medical Vision-Language Models
Shakeri, Fereshteh (TS Montreal), Ben Ayed, Ismail (ÉTS Montréal)
CodeDomain AdaptationRepresentation LearningTransformerVision Language ModelContrastive LearningMultimodalityBiomedical DataBenchmark
🎯 What it does: Evaluate the adaptation of medical visual language models with a small number of samples, establish benchmarks, and compare various adaptation strategies.
🎯 What it does: An information-based non-parametric classifier (INC) is proposed, which maintains local patch features on a small number of labeled WSIs and uses mask labels for non-parametric similarity matching to achieve few-shot classification of lymph node metastasis.
🎯 What it does: By combining a pre-trained general segmentation model with active few-shot labeling, the FM-ABS learning framework is proposed, which allows for training a dedicated segmentation model with only three labeled slices in 3D medical image segmentation.
🎯 What it does: This paper proposes a framework FM-OSD for single image one-shot anatomical landmark point detection using a visual foundation model, which can achieve high-precision landmark localization with just one template image.
🎯 What it does: This paper proposes a deep learning model based on Parallel Hyperplanes for predicting the risk of developing dry age-related macular degeneration (dAMD) and the cumulative distribution function (CDF) over continuous time from retinal OCT scans; it also utilizes unlabeled longitudinal image pairs for unsupervised fine-tuning to adapt to domain shifts from different scanners.
FRCNet: Frequency and Region Consistency for Semi-supervised Medical Image Segmentation
He, Along (Nankai University), Fu, Huazhu (Agency for Science, Technology and Research (A*STAR))
CodeSegmentationTransformerImageBiomedical Data
🎯 What it does: Proposes FRCNet, which enhances semi-supervised medical image segmentation performance using frequency domain consistency and multi-scale region similarity.
Free-SurGS: SfM-Free 3D Gaussian Splatting for Surgical Scene Reconstruction
Guo, Jiaxin (Chinese University of Hong Kong), Liu, Yun-hui (Hong Kong Center for Logistics Robotics)
CodeDepth EstimationOptimizationGaussian SplattingSimultaneous Localization and MappingOptical FlowVideo
🎯 What it does: This paper proposes Free‑SurGS, the first SfM-free 3D Gaussian Splatting method, which enables rapid reconstruction and real-time rendering of surgical scenes using monocular endoscopic video.
🎯 What it does: This paper proposes the Pixel2Cancer method, which generates multi-stage, cross-organ synthetic tumors in computed tomography images based on three general rules of cellular automata (growth, invasion, death) without the need for manual annotation.
From Static to Dynamic Diagnostics: Boosting Medical Image Analysis via Motion-Informed Generative Videos
Li, Wuyang (Chinese University of Hong Kong), Yuan, Yixuan (Southern University of Science and Technology)
CodeClassificationGenerationData SynthesisKnowledge DistillationDiffusion modelImageVideoBiomedical Data
🎯 What it does: Utilizing video generation models to convert static medical images into medical videos with motion information, and jointly using image and video features in model training to enhance the performance of semi-supervised medical image classification.
🎯 What it does: A framework for fMRI-Text Synergistic Prompt Learning (fTSPL) is proposed, which utilizes a pre-trained Vision-Language model to automatically generate instance-level text and construct multimodal functional connectivity graphs, thereby enhancing brain function analysis based on graph neural networks.
🎯 What it does: Train a shared backbone network under the federated learning framework, and learn a segmentation head separately for each client. After estimating uncertainty using MC Dropout, obtain the global segmentation result through uncertainty-weighted averaging.
🎯 What it does: A model based on autoregressive GAN (Fundus2Video) is proposed, capable of synthesizing dynamic fluorescein angiography (FFA) videos from a single color fundus photograph (CF).
Gaussian Pancakes: Geometrically-Regularized 3D Gaussian Splatting for Realistic Endoscopic Reconstruction
Bonilla, Sierra (Wellcome/EPSRC Centre for Interventional and Surgical Sciences), Bano, Sophia (Wellcome/EPSRC Centre for Interventional and Surgical Sciences)
CodeRestorationGenerationDepth EstimationRecurrent Neural NetworkGaussian SplattingSimultaneous Localization and MappingVideo
🎯 What it does: The paper introduces 'Gaussian Pancakes', which combines 3D Gaussian Splatting with an RNN-based SLAM system to achieve real-time and accurate 3D reconstruction and view synthesis in endoscopic videos.
Gaze-DETR: Using Expert Gaze to Reduce False Positives in Vulvovaginal Candidiasis Screening
Kong, Yan (Nanjing University), Wang, Qian (United Imaging Intelligence)
CodeObject DetectionTransformerImage
🎯 What it does: This paper proposes a detection method using expert eye movement data to reduce false positives in vaginal candidiasis screening—Gaze-DETR;
Gaze-directed Vision GNN for Mitigating Shortcut Learning in Medical Image
Wu, Shaoxuan (Northwest University), Feng, Jun (Northwest University)
CodeClassificationConvolutional Neural NetworkGraph Neural NetworkImageBiomedical Data
🎯 What it does: This paper proposes a Vision GNN (GD-ViG) based on retinal eye movements, which guides the network to focus on lesion areas by generating pupil maps, thereby reducing shortcut learning in medical imaging.
GCAN: Generative Counterfactual Attention-guided Network for Explainable Cognitive Decline Diagnostics based on fMRI Functional Connectivity
Shen, Xiongri (Harbin Institute of Technology (Shenzhen)), Zhang, Zhiguo (Harbin Institute of Technology)
CodeClassificationExplainability and InterpretabilityTransformerGenerative Adversarial NetworkBiomedical DataMagnetic Resonance ImagingAlzheimer's Disease
🎯 What it does: A Generative Adversarial Causal Attention Network (GCAN) is proposed to explain and enhance the diagnosis of mild cognitive impairment and subjective cognitive decline based on fMRI functional connectivity.
🎯 What it does: This paper proposes the GEM network, which implements context-based gaze estimation for medical images by utilizing fine-grained alignment between images and reports, as well as visual behavior graph matching, to predict the gaze points of radiologists on chest X-rays.