🎯 What it does: This paper proposes a shape prior-based network (SHAN) that learns the elliptical distribution of thyroid nodules through global and neighborhood affine modules, constructing a domain-invariant latent feature space to achieve cross-center ultrasound image nodule segmentation.
🎯 What it does: This study investigates the shortcut learning phenomenon in medical image segmentation and demonstrates two common shortcuts: annotation markers in ultrasound images and center cropping + zero padding in skin lesion segmentation.
CodeSegmentationConvolutional Neural NetworkLarge Language ModelImageBiomedical DataMagnetic Resonance Imaging
🎯 What it does: Using simple text prompts to achieve weakly supervised medical image segmentation, the SimTxtSeg framework is constructed, which first generates visual prompts from text to obtain pseudo-labels, and then trains the segmentation model.
🎯 What it does: This paper addresses the single-source domain generalization problem and proposes a spectrum-based Lipschitz regularization method (LRFS) for medical image segmentation.
🎯 What it does: A physics-based domain randomization method called SinoSynth is proposed, which can synthesize various CBCT artifacts on CT images and automatically generate aligned CBCT-CT paired training data.
🎯 What it does: We propose and implement SlicerTMS, an open-source system capable of rapidly predicting and visualizing transcranial magnetic stimulation electric fields in a real-time brain navigation environment using deep learning.
SlideGCD: Slide-based Graph Collaborative Training with Knowledge Distillation for Whole Slide Image Classification
Shu, Tong (Hefei University of Technology), Zheng, Yushan (Beihang University)
CodeClassificationKnowledge DistillationGraph Neural NetworkImageBiomedical Data
🎯 What it does: The SlideGCD framework is proposed, treating the entire pathological slide as a graph node, constructing a large-scale slide-level graph, and enhancing WSI classification performance through knowledge distillation.
🎯 What it does: A spatial distribution-guided diffusion model is proposed for generating cell layouts and pathological images, thereby enhancing cell detection performance.
Spatial Transcriptomics Analysis of Zero-shot Gene Expression Prediction
Yang, Yan (Australian National University), Stone, Eric (Australian National University)
CodeGraph Neural NetworkTransformerLarge Language ModelPrompt EngineeringBiomedical Data
🎯 What it does: Proposes a Semantic Guided Network (SGN) to achieve zero-shot gene expression prediction based on spatial transcriptomics image windows.
CodeAnomaly DetectionGenerative Adversarial NetworkImageBiomedical Data
🎯 What it does: A spatial attention-based Generative Adversarial Network (SAGAN) is proposed, which generates healthy images to restore abnormal regions and uses image differences as anomaly scores for semi-supervised medical image anomaly detection.
🎯 What it does: A SC-Net model aimed at the scarcity of coronary CT angiography data is proposed, achieving efficient automatic diagnosis of coronary artery diseases.
🎯 What it does: A conditional generative model based on neural implicit distance fields (SDF) is proposed, which can complete the entire cardiac cycle given static atrial images and clinical demographic information, and generate new dynamic atrial sequences based on clinical information.
🎯 What it does: A spatiotemporal graph neural network called PerfGAT is proposed to predict the IDH mutation status of brain gliomas using dynamic contrast-enhanced MRI.
🎯 What it does: This paper proposes two self-supervised spatiotemporal representation learning methods, namely c SimCLR based on single-slice contrastive learning and TVRL which combines frame-level feature prediction, to improve the representation of medical image time series.
🎯 What it does: This paper presents SDSeg, a medical image segmentation framework based on Stable Diffusion, which can generate segmentation results in a single-step reverse process, eliminating the need for traditional multi-step sampling and multiple sample averaging.
Stealing Knowledge from Pre-trained Language Models for Federated Classifier Debiasing
Zhu, Meilu (City University of Hong Kong), Yuan, Yixuan (Southern University of Science and Technology)
CodeClassificationFederated LearningConvolutional Neural NetworkLarge Language ModelPrompt EngineeringImageTextBiomedical Data
🎯 What it does: The FedCB framework is proposed, which constructs a fixed classifier by extracting text embeddings from pre-trained language models and aligns image features with text distributions during the federated learning process, thereby addressing the classifier bias problem caused by heterogeneous data.
🎯 What it does: In radiation therapy, the subgroup risk control prediction set (SG-RCPS) provides a reliable uncertainty interval for deep learning dose prediction.
Subject-Adaptive Transfer Learning Using Resting State EEG Signals for Cross-Subject EEG Motor Imagery Classification
An, Sion (Daegu Gyeongbuk Institute of Science and Technology), Park, Sang Hyun (DGIST)
CodeClassificationDomain AdaptationConvolutional Neural NetworkTransformerTime SeriesBiomedical Data
🎯 What it does: An adaptive transfer learning framework called ResTL is proposed, based on resting-state EEG signals, for motor imagery (MI) classification in cross-subject brain-computer interfaces.
🎯 What it does: A network named SFNet has been developed, utilizing ultra-low field (0.064T) infant MRI to generate dual-channel paired data through latent diffusion, producing high-quality images equivalent to high-field (≥1.5T) MRI;
Surgformer: Surgical Transformer with Hierarchical Temporal Attention for Surgical Phase Recognition
Yang, Shu (Hong Kong University of Science and Technology), Chen, Hao (Harvard Medical School)
CodeRecognitionTransformerVideo
🎯 What it does: This paper proposes an end-to-end Surgformer model that achieves surgical phase recognition through sparse frame sequences using separated spatiotemporal attention.
Survival analysis of histopathological image based on a pretrained hypergraph model of spatial transcriptomics data
Cai, Shangyan (Beijing Normal University-Hong Kong Baptist University United International College), Su, Weifeng (Beijing Normal-Hong Kong Baptist University)
CodeGraph Neural NetworkImageMultimodalityBiomedical Data
🎯 What it does: This paper proposes a multimodal hypergraph neural network (MHNN-surv) that integrates histopathological images with spatial gene expression information through a pretrained spatial transcriptomics model to perform survival analysis for breast cancer.
SurvRNC: Learning Ordered Representations for Survival Prediction using Rank-N-Contrast
Saeed, Numan (Mohamed bin Zayed University of Artificial Intelligence), Yaqub, Mohammad (Mohamed bin Zayed University of Artificial Intelligence)
CodeRepresentation LearningConvolutional Neural NetworkContrastive LearningMultimodalityBiomedical DataComputed TomographyPositron Emission TomographyElectronic Health Records
🎯 What it does: A new SurvRNC loss function is proposed and implemented on multimodal head and neck cancer data to learn latent representations ordered by survival time, and it is integrated into deep survival models (DeepMTLR, DeepHit).
Swin-UMamba: Mamba-based UNet with ImageNet-based pretraining
Liu, Jiarun (Paul C. Lauterbur Research Center for Biomedical Imaging, Shenzhen Institute of Advanced Technology, Chinese Academy of Sciences), Wang, Shanshan (Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences)
🎯 What it does: A Mamba-based UNet network called Swin-UMamba is proposed, specifically for medical image segmentation, and its performance is enhanced through ImageNet pre-training.
🎯 What it does: A Symmetry-Aware Cross-Attention (SACA) module and Symmetry-Aware Head (SAH) based on the symmetrical features of the left and right hemispheres are proposed, enhancing brain imaging classification and segmentation performance through symmetry pre-training.
Symptom Disentanglement in Chest X-ray Images for Fine-Grained Progression Learning
Zhu, Ye (Hong Kong Baptist University), Yuen, Pong C. (Chinese University of Hong Kong)
CodeClassificationRecognitionTransformerImageBiomedical Data
🎯 What it does: A chest X-ray progression learning framework based on symptom disentanglement is proposed, utilizing a Symptom Disentangler to extract symptom features and capturing symptom-level progression through a Symptom Progression Learner.
SynCellFactory: Generative Data Augmentation for Cell Tracking
Sturm, Moritz (Heidelberg University), Hamprecht, Fred A. (Heidelberg University)
CodeObject TrackingGenerationData SynthesisDiffusion modelVideoBiomedical Data
🎯 What it does: Construct a generative data augmentation pipeline called SynCellFactory based on ControlNet, designed for the automatic generation of cell videos with realistic segmentation and tracking labels.
🎯 What it does: This paper proposes a synchronous image-label diffusion model, replacing traditional isotropic Gaussian noise with anisotropic noise for the automatic segmentation of stroke lesions in non-contrast CT images.
🎯 What it does: Using cardiac magnetic resonance (CMR) videos and tabular features, we propose the TabMixer module to construct a deep learning framework for non-invasive measurement of mean pulmonary artery pressure (mPAP).
Tackling Data Heterogeneity in Federated Learning via Loss Decomposition
Zeng, Shuang (The University of Hong Kong), Qu, Liangqiong (The University of Hong Kong)
CodeFederated LearningImageBiomedical Data
🎯 What it does: This paper proposes to decompose the global loss of federated learning into local loss, distribution shift loss, and aggregation loss, and based on this, designs two methods: Margin control regularization and main gradient aggregation, to jointly reduce the three types of losses and achieve more robust federated learning.
🎯 What it does: A method for generating temporal-aware brain MRI progress based on diffusion models is proposed and implemented, predicting future MRIs by learning the intensity differences between baseline and follow-up scans.
TaGAT: Topology-Aware Graph Attention Network For Multi-modal Retinal Image Fusion
Tian, Xin (University of Bristol), Achim, Alin (University of Bristol)
CodeImage TranslationRestorationData SynthesisGraph Neural NetworkTransformerImageMultimodalityBiomedical Data
🎯 What it does: A multi-modal retinal image fusion framework called TaGAT is designed and implemented, which combines spatial features with vascular topological structures through a topology-aware graph attention network to achieve high-quality fused image generation.
🎯 What it does: A tail class enhancement representation learning framework (TERL) based on multi-task learning, instance-level contrastive learning, and prototype semantic enhancement is designed to improve the recognition performance of tail classes in surgical video triplet recognition.
TAKT: Target-Aware Knowledge Transfer for Whole Slide Image Classification
Xiong, Conghao (Chinese University of Hong Kong), King, Irwin (Chinese University of Hong Kong)
CodeClassificationDomain AdaptationKnowledge DistillationSupervised Fine-TuningImageBiomedical Data
🎯 What it does: This paper proposes a target-aware knowledge transfer framework (TAKT) for whole-slide image classification, achieving knowledge transfer through a teacher-student paradigm that combines target-aware data augmentation and feature alignment.
🎯 What it does: A self-supervised learning framework TE-SSL is proposed, which combines event labels and time difference weights for imaging feature extraction and survival analysis of Alzheimer's disease progression.
🎯 What it does: The TeethDreamer framework is proposed, which generates multi-view color images and normal maps from five intraoral photos, and obtains high-quality 3D tooth models through neural surface reconstruction.
TextPolyp: Point-supervised Polyp Segmentation with Text Cues
Zhao, Yiming (Nanjing University of Science and Technology), Zhou, Tao (Nanjing University of Science and Technology)
CodeSegmentationImage
🎯 What it does: A weakly supervised polyp segmentation method based on SAM and text prompts has been developed, which can generate high-quality segmentation results with only point annotations.
🎯 What it does: A reinforcement learning-based active sampling framework for MRI k-space is proposed, which directly performs lesion diagnosis on un-reconstructed k-space data.
This actually looks like that: Proto-BagNets for local and global interpretability-by-design
Djoumessi, Kerol (University of Tbingen), Koch, Lisa (University of Tbingen)
CodeClassificationExplainability and InterpretabilityConvolutional Neural NetworkImageBiomedical DataMagnetic Resonance Imaging
🎯 What it does: This paper proposes an interpretable Proto-BagNet model for detecting drusen, a lesion related to age-related macular degeneration, in retinal optical coherence tomography (OCT) images, providing local and global interpretable prototype images during the inference phase.
ThyGraph: A Graph-Based Approach for Thyroid Nodule Diagnosis from Ultrasound Studies
Radhachandran, Ashwath (UCLA), Speier, William (UCLA)
CodeClassificationExplainability and InterpretabilityGraph Neural NetworkImageBiomedical DataUltrasound
🎯 What it does: A ThyGraph method based on graph convolutional networks is proposed, which constructs patient-level graphs using multi-view ultrasound images to predict the malignancy of thyroid nodules.
🎯 What it does: Design and implement a Temporal Latent Residual Network (TLRN) that predicts continuous deformation fields of time series images by recursively learning the residual function in the latent velocity field space, achieving high-precision time series registration.
🎯 What it does: A method based on the Topological Cycle Graph Attention Network (CycGAT) is proposed to identify functional backbones from brain functional connectivity graphs and eliminate redundant edges.
Topological SLAM in colonoscopies leveraging deep features and topological priors
Morlana, Javier (University of Zaragoza), Montiel, José M. M. (Universidad de Zaragoza)
CodeTransformerSimultaneous Localization and MappingImageVideoBiomedical Data
🎯 What it does: This paper presents ColonSLAM, a system that combines classical multi-map metric SLAM with deep features and topological priors to construct a topological map of the entire colon.
Towards a text-based quantitative and explainable histopathology image analysis
Nguyen, Anh Tien (Korea University), Kwak, Jin Tae (Korea University)
CodeClassificationRetrievalExplainability and InterpretabilityTransformerVision Language ModelImageTextMultimodalityBiomedical Data
🎯 What it does: The TQx framework is proposed, utilizing a pre-trained vision-language model to generate interpretable text features through image-text retrieval, achieving quantitative analysis of pathological images.
Qian, Kui (University of California, San Diego), Freund, Yoav (University of California, San Diego)
CodeClassificationExplainability and InterpretabilityDiffusion modelBiomedical Data
🎯 What it does: Utilizing cell shape features for interpretable automatic identification of brain structures, constructing region features based on cell distribution and using XGBoost for classification;
Towards Graph Neural Networks with Domain-Generalizable Explainability for fMRI-Based Brain Disorder Diagnosis
Qiu, Xinmei, Ma, Jianhua (Pazhou Lab)
CodeDomain AdaptationExplainability and InterpretabilityMeta LearningGraph Neural NetworkBiomedical DataMagnetic Resonance Imaging
🎯 What it does: This paper proposes a graph neural network called XG-GNN that can simultaneously achieve interpretability and domain generalization for multi-center fMRI brain disease diagnosis.
🎯 What it does: This paper studies the generation of 3D ultrasound volumes from preoperative 2D ultrasound scans through registration, and trains a personalized Attention U-Net model based on this volume to achieve real-time identification of intrahepatic vessels during laparoscopic surgery.
CodeSegmentationConvolutional Neural NetworkPrompt EngineeringContrastive LearningImageBiomedical Data
🎯 What it does: A text prompt-assisted Segment Anything Model (TP-DRSeg) framework is proposed for the segmentation of diabetic retinopathy (DR) lesions.
🎯 What it does: Using pure Vision Transformer (ViT) for Alzheimer's classification with limited brain MRI data, this study systematically explores the effects of various training strategies such as self-supervised pre-training, knowledge distillation, and data augmentation.
🎯 What it does: A training-free video diffusion model (Free-Echo) is proposed, which can generate realistic cardiac ultrasound videos based solely on a single frame of cardiac ultrasound segmentation images.
🎯 What it does: Transfer the relative monocular depth model trained on natural images to endoscopic images, significantly improving depth estimation accuracy through time-consistent self-supervision.
Transforming Surgical Interventions with Embodied Intelligence for Ultrasound Robotics
Xu, Huan (Centre for Artificial Intelligence and Robotics, HKISI-CAS), Liu, Hongbin (Centre for Artificial Intelligence and Robotics, Hong Kong Institute of Science Innovation, Chinese Academy of Sciences)
CodeRobotic IntelligenceTransformerLarge Language ModelTextBiomedical DataUltrasoundRetrieval-Augmented Generation
🎯 What it does: This paper proposes an embedded intelligent system based on large language models and domain knowledge enhancement, achieving closed-loop control for precise motion planning and dynamic execution of ultrasound robots from voice commands.
🎯 What it does: Trexplorer is a recursive DETR model designed to track the centerlines of tree structures in 3D medical images, ensuring topological correctness without the need for post-processing.
🎯 What it does: This paper proposes the TP-Mamba adapter, which migrates the Segment Anything Model (SAM) to 3D medical image segmentation tasks, achieving efficient adaptation in terms of parameters and data.
TSBP: Improving Object Detection in Histology Images via Test-time Self-guided Bounding-box Propagation
Yang, Tingting (Nanjing University of Science and Technology), Zhang, Yizhe (Nanjing University of Science and Technology)
CodeObject DetectionConvolutional Neural NetworkDiffusion modelImageBiomedical Data
🎯 What it does: A method called Test-time Self-guided Bounding Box Propagation (TSBP) is proposed, which allows high-confidence detection boxes to influence low-confidence boxes, thereby improving the recall and accuracy of object detection in histology images without the need for additional labeled samples.
🎯 What it does: A latent dynamic diffusion model (LDDM) is proposed, which can synthesize playable ultrasound videos from a single ultrasound image and use these synthesized videos for training video classification models, thereby alleviating the problem of scarcity of clinical ultrasound video data.
🎯 What it does: A framework for cardiac image segmentation is proposed, which utilizes a diffusion generative model to synthesize diverse ultrasound images and optimizes through meta-learning spatial weighting, achieving domain generalization.
🎯 What it does: A general medical image registration foundational model called uniGradICON is proposed, which can achieve fast and accurate registration on images from various anatomical regions, modalities, and sources;
CodeClassificationDomain AdaptationConvolutional Neural NetworkAuto EncoderImageBiomedical Data
🎯 What it does: A unified framework is proposed to utilize unlabeled data from unknown categories and unknown domains for general semi-supervised learning in medical image classification.
🎯 What it does: A general topology refinement post-processing module is proposed, which enhances the topological accuracy of medical image segmentation through a plug-and-play network trained with topological perturbation masks synthesized by orthogonal polynomials.
🎯 What it does: UrFound is proposed, a universal retinal foundation model capable of simultaneously processing CFP and OCT images, and learns cross-modal universal representations through knowledge-guided masked modeling.
VDPF: Enhancing DVT Staging Performance Using a Global-Local Feature Fusion Network
Xie, Xiaotong (Affiliated Panyu Central Hospital Guangzhou Medical University), Huang, Yi (Affiliated Panyu Central Hospital Guangzhou Medical University)
🎯 What it does: This paper proposes a deep vein thrombosis (DVT) staging system based on black blood magnetic resonance imaging (BTI), constructing the VDPF framework and achieving joint analysis of global images and local lesion images through a global-local feature fusion module.
🎯 What it does: Proposes the VertFound framework, which integrates the semantic understanding of CLIP and the spatial localization of SAM to achieve fine-grained vertebral classification in spinal images.
🎯 What it does: Utilizing neural fields and a time-conditioned recurrent module to predict the morphological changes of vestibular schwannomas over time.
🎯 What it does: Persistent homology is applied to the functional connectivity networks generated by fMRI in the Human Connectome Project to extract the volume-optimal persistent cycles and calculate their persistent centrality. Subsequently, these high-order topological features are aligned and compared with the aperiodic power changes of MEG signals in the theta-alpha (4-12 Hz) range, exploring their spatial distribution and variations under different attention tasks.
🎯 What it does: A three-dimensional PET/MR denoising model based on Conditional Residual Diffusion (CSRD) is proposed, utilizing 3D patch training for efficient denoising.
🎯 What it does: Detect and segment cerebral hemorrhage in non-contrast head CT scans and generate a 3D voxel scene map, with the model simultaneously learning the relationship between hemorrhage and brain structures to provide clinical decision support.
🎯 What it does: A CoSeg framework is proposed, utilizing weakly supervised brain tissue segmentation data to train a temporal attention network, achieving differential homeomorphic deformation of the initial cortical mesh for rapid and accurate cortical surface reconstruction.
🎯 What it does: A weakly supervised tooth instance segmentation network WS-TIS is proposed, which completes the segmentation of tooth instances in 3D dental models using multi-label learning with only 50% point-level annotations and thematic labels.
Weakly-supervised Medical Image Segmentation with Gaze Annotations
Zhong, Yuan (Chinese University of Hong Kong), Dou, Qi (Huazhong University of Science and Technology)
CodeSegmentationConvolutional Neural NetworkImageBiomedical Data
🎯 What it does: This paper proposes the use of eye-tracking fixation data as weak supervision to achieve medical image segmentation through multi-layer networks and consistency regularization.
🎯 What it does: In the task of tooth point cloud segmentation with extremely sparse labels (only one point labeled per tooth), the SAMTooth framework is proposed, which utilizes the prompt capability of SAM to automatically generate point prompts and projects the 2D masks generated by SAM back into 3D space, achieving high-precision segmentation through contrastive learning.
🎯 What it does: Using a self-supervised mask reconstruction framework, we perform encoding learning on multi-view 2D+T CMR images and generate a unified 3D+T representation for cardiac phenotype prediction and whole heart segmentation.
WiNet: Wavelet-based Incremental Learning for Efficient Medical Image Registration
Cheng, Xinxing (University of Birmingham), Duan, Jinming (University of Birmingham)
CodeSegmentationOptimizationExplainability and InterpretabilityComputational EfficiencyConvolutional Neural NetworkImageBiomedical DataMagnetic Resonance Imaging
🎯 What it does: This paper proposes an incremental learning network called WiNet based on wavelet transform for efficient and interpretable 3D medical image registration.
WsiCaption: Multiple Instance Generation of Pathology Reports for Gigapixel Whole-Slide Images
Chen, Pingyi (Zhejiang University), Yang, Lin (Westlake University)
CodeGenerationTransformerLarge Language ModelImageText
🎯 What it does: Collected and constructed nearly 10,000 pairs of high-quality whole slide images and pathology reports (PathText), and proposed a multi-instance generation framework (MI-Gen) to achieve gigapixel-level automatic generation of pathology reports.
🎯 What it does: A weakly supervised spherical age decoupling network (WSSADN) is proposed, which extracts disease-related features for diagnosing developmental disorders by removing age information related to normal development in brain structures.
🎯 What it does: A framework for unlabeled X-ray vascular segmentation, XA-Sim2Real, is proposed, achieving high-quality segmentation through the alignment of simulated images and adaptive representations.
XCoOp: Explainable Prompt Learning for Computer-Aided Diagnosis via Concept-guided Context Optimization
Bie, Yequan (Hong Kong University of Science and Technology), Chen, Hao (Harvard Medical School)
CodeOptimizationExplainability and InterpretabilityTransformerPrompt EngineeringVision Language ModelContrastive LearningImageBiomedical DataMagnetic Resonance Imaging
🎯 What it does: This paper proposes an explainable prompt learning framework called XCoOp, which achieves interpretability and high performance in medical image diagnosis through multi-granularity soft and hard prompt alignment.
XTranPrune: eXplainability-aware Transformer Pruning for Bias Mitigation in Dermatological Disease Classification
Ghadiri, Ali (University of New South Wales), Song, Yang (University of New South Wales)
CodeClassificationExplainability and InterpretabilityTransformerImage
🎯 What it does: This paper proposes XTranPrune, an interpretable method for pruning visual Transformers (DeiT) to mitigate gender/skin color bias in skin disease classification.