🎯 What it does: A framework for accelerated cardiac T1 mapping based on physics-informed neural ODEs is proposed, achieving high-precision T1 estimation under sparse sampling.
Physics-Informed Neural Operators for Tissue Elasticity Reconstruction
Kim, Youjin (Chung-Ang University), Kwon, Junseok (Chung-Ang University)
CodeBiomedical DataMagnetic Resonance ImagingPhysics Related
🎯 What it does: This paper proposes a magnetic resonance elastography (MRE) elastic reconstruction framework based on a physics-informed neural operator (MRE-Hyper), which can learn the mapping from wave images to elastic fields in one go;
🎯 What it does: Using implicit neural representations (INR) for voxel-level reversible two-compartment kinetic model (TCKM) parameter estimation of dynamic [18F]FDG PET data, constructing continuous spatiotemporal parameter mappings;
🎯 What it does: A pluggable plugin framework called PLUS is proposed to enhance the existing 3D segmentation models for liver lesion diagnosis in non-contrast CT scans.
PolyMamba: Spatial-prior Guided Mamba for Polyp Segmentation with High-Frequency Enhancement
Fu, Renyu, Liu, Xinwang (National University Of Defense Technology)
CodeSegmentationTransformerImageBiomedical Data
🎯 What it does: Proposes the PolyMamba framework, which integrates spatial priors generated by Transformer into the Mamba state space model, and designs a dual-gate frequency domain enhancement module to improve polyp segmentation accuracy.
🎯 What it does: This paper presents a large-scale expert-annotated Periapical Radiograph Dataset (PRAD-10K) and a multi-scale Wavelet Convolution Network (PRNet) specifically designed for this dataset, aimed at pixel-level segmentation and classification of dental PR images.
🎯 What it does: A reinforcement learning-based active video collection framework, PRECISE-AS, was designed for personalized selection of cardiac ultrasound videos to diagnose aortic stenosis.
🎯 What it does: A three-stream deep learning framework is proposed, which combines T1-MRI, ROI annotations, and ONFH grading information to perform survival analysis prediction of the risk of femoral head collapse in the early stages of avascular necrosis using label tokenization and temporal ordering representation.
🎯 What it does: A framework is constructed using longitudinal dynamic MRI images to predict pathological complete response (pCR)/treatment response in breast cancer patients after neoadjuvant chemotherapy (or radiotherapy).
CodeRepresentation LearningTransformerAuto EncoderImageBiomedical Data
🎯 What it does: An integrated self-supervised retinal foundation model PRETI was constructed, incorporating patient metadata (age, gender) for efficient learning of representations from color retinal photographs;
🎯 What it does: Pre-trained a Masked Autoencoder (MAE) on the COPDGene dataset for chronic pneumonia and fine-tuned the pre-trained model for the classification of lung nodule malignancy on the NLST dataset.
🎯 What it does: This paper proposes a hierarchical regional graph neural network (PRGNN) that extracts multi-scale features through 3D CNN and utilizes brain region ROIs to achieve node embedding for disease classification of PET images.
Prior-guided Prototype Aggregation Learning for Alzheimer’s Disease Diagnosis
Diao, Yueqin (South China University of Technology), Xu, Yanwu (South China University of Technology)
CodeClassificationConvolutional Neural NetworkLarge Language ModelContrastive LearningImageBiomedical DataMagnetic Resonance ImagingAlzheimer's Disease
🎯 What it does: This paper proposes a prototype aggregation learning framework (PPAL) guided by prior knowledge for the diagnosis of Alzheimer's disease.
🎯 What it does: Proposes a sparse information probabilistic integration of radiomics and pathology to predict overall survival in kidney cancer patients by aggregating spatial features using graph neural networks;
Prolog-Driven Rule-Based Diagnostics with Large Language Models for Precise Clinical Decision Support
Tan, Xiaoyu (Shanghai University of Engineering Science), Qiu, Xihe (Shanghai University of Engineering Science)
CodeTransformerLarge Language ModelPrompt EngineeringTextBiomedical DataChain-of-Thought
🎯 What it does: A clinical decision support system called ProCDS has been developed, which combines Prolog rule reasoning with large language models (LLM) to reduce errors caused by LLM hallucinations and provide transparent and accurate diagnostic suggestions.
🎯 What it does: We propose Prompt-DAS, a prompt-based Transformer that can perform multi-instance segmentation using sparse point prompts or no prompts in unsupervised, weakly supervised, and interactive domain adaptation tasks for electron microscope image segmentation.
🎯 What it does: This paper proposes PromptReg, which achieves the generalization of a single model in multi-domain medical image registration tasks through prompt learning.
Prototype-Based Multiple Instance Learning for Gigapixel Whole Slide Image Classification
Sun, Susu (University of Tübingen), Baumgartner, Christian F. (University of Tübingen)
CodeClassificationExplainability and InterpretabilityAuto EncoderImageBiomedical Data
🎯 What it does: ProtoMIL has been developed, a multi-instance learning model based on sparse autoencoders that automatically discovers interpretable concepts for whole slide image classification and supports human intervention.
Prototype-Guided and Lightweight Adapters for Inherent Interpretation and Generalisation in Federated Learning
Ofosu Mensah, Samuel (University of Tübingen), Berens, Philipp (University of Tübingen)
CodeFederated LearningExplainability and InterpretabilityConvolutional Neural NetworkImage
🎯 What it does: A framework is proposed that only transmits adapter and prototype parameters in federated learning to address statistical heterogeneity and provide model interpretability.
Prototype-Guided Cross-Modal Knowledge Enhancement for Adaptive Survival Prediction
Liu, Fengchun (Harbin Institute of Technology Shenzhen), Zhang, Yongbing (Harbin Institute of Technology Shenzhen)
CodeClassificationContrastive LearningMultimodalityBiomedical Data
🎯 What it does: The Prototype-Guided Cross-Modal Knowledge Enhancement (ProSurv) framework is proposed to achieve adaptive survival prediction with unpaired multimodal data.
🎯 What it does: This paper constructs the PSAT framework and systematically evaluates pediatric CT segmentation strategies based on adult augmentation and transfer learning.
🎯 What it does: This paper proposes a ViT model called PTCMIL based on visual prompt word clustering for feature aggregation and prediction in multi-instance learning on whole slide images.
🎯 What it does: A Q-space guided collaborative attention translation network (Q-CATN) is proposed to achieve flexible synthesis of multi-shell high angular resolution diffusion imaging (MS-HARDI) based on structural MRI.
Query-Level Alignment for End-to-End Lesion Detection with Human Gaze
Kong, Yan (Nanjing University), Shan, Caifeng (Nanjing University)
CodeObject DetectionExplainability and InterpretabilityTransformerGaussian SplattingImage
🎯 What it does: The study integrates clinical eye movement data into a Transformer-based medical lesion detection model, proposing GAA-DETR to achieve query-level alignment, enhancing detection accuracy and interpretability.
R1Seg-3D: Rethinking Reasoning Segmentation for Medical 3D CTs
Hao, Qin (Xinjiang University), Zhang, Lei (University of Exeter)
CodeSegmentationTransformerLarge Language ModelSupervised Fine-TuningImageBiomedical DataComputed Tomography
🎯 What it does: A three-dimensional medical CT inference segmentation framework R1Seg-3D is proposed, which unifies visual, textual reasoning, and mask decoding in the same latent space.
CodeGenerationRetrievalConvolutional Neural NetworkTransformerLarge Language ModelVision Language ModelContrastive LearningTextMultimodalityBiomedical DataElectronic Health RecordsRetrieval-Augmented Generation
🎯 What it does: The RadAlign framework has been developed to achieve more accurate and interpretable radiology report generation by aligning visual features with medical diagnostic concepts.
Radar-Based Imaging for Sign Language Recognition in Medical Communication
Mineo, Raffaele (University of Catania), Palazzo, Simone (University of Catania)
CodeRecognitionSafty and PrivacyConvolutional Neural NetworkTransformerAuto EncoderMultimodalityBiomedical Data
🎯 What it does: A privacy-preserving medical sign language recognition framework based on 60 GHz millimeter-wave radar is proposed, achieving automatic recognition of 126 medical terms and letters in Italian sign language.
🎯 What it does: A hierarchical Transformer model named RadioFormer is proposed for tumor classification on multi-sequence MRI data. The model is divided into three layers: single-sequence slice feature extraction → multi-sequence slice information aggregation → inter-slice (volume) information aggregation, simulating the diagnostic process of radiologists and enabling efficient learning on small-scale datasets.
🎯 What it does: This paper proposes the RadiomicsRetrieval framework, which combines Radiomics features of 3D medical images with deep learning embeddings to achieve tumor-level retrieval based on image, location, or partial features.
RAPTOR: Generative AI for Parsing Colorectal Cancer Referrals to Streamline Faster Diagnostic Standard Pathways
Abioye, Sofiat (Birmingham City University), Bilal, Muhammad (Birmingham City University)
CodeTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextElectronic Health Records
🎯 What it does: This paper develops and evaluates a generative AI-based system called RAPTOR, designed to automatically extract structured information from emergency referral forms for colorectal cancer.
🎯 What it does: A recursive reasoning and representation decoupled multimodal large deformation registration network (RDMR) is proposed to address the challenges of strong intensity differences and large-scale tissue deformations between different modalities.
🎯 What it does: This study addresses the low-resolution issue of clinical CT images by using a deep learning super-resolution model to reconstruct the proximal femur. An automated point cloud registration method is employed to extract the ROI and calculate bone microstructure indices (BV/TV, Tb.Th, Tb.Sp, Tb.N) to assess the model's potential in predicting bone strength.
🎯 What it does: This paper proposes a framework for mediastinal lymph node segmentation based on incomplete supervision, ReCo-I2P, which utilizes extremely sparse orthogonal partial instance annotations (oPIA) for training and achieves high-quality segmentation with only a small amount of annotations.
Reconsidering Explicit Longitudinal Mammography Alignment for Enhanced Breast Cancer Risk Prediction
Thrun, Solveig (UiT Arctic University of Norway), Kampffmeyer, Michael (UiT Arctic University of Norway)
CodeClassificationConvolutional Neural NetworkTransformerImageBiomedical Data
🎯 What it does: This paper evaluates explicit alignment methods for longitudinal mammography images, proposing and implementing the first deep learning-based mammography image registration model, MammoRegNet, and comparing the effects of image-level and feature-level alignment in risk prediction tasks.
🎯 What it does: This paper proposes a method to reconstruct a 3D interaction model of a doctor's hand and medical instruments from a single 2D medical scene image, including steps such as hand initialization, instrument initialization, and a contact point-based interaction module (CPCI).
🎯 What it does: Developed RedDino - a foundational model for red blood cell analysis based on self-supervised learning, aimed at efficiently and accurately identifying red blood cell morphology;
CodeClassificationExplainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringVision Language ModelImageMultimodalityBiomedical DataRetrieval-Augmented Generation
🎯 What it does: A visual-language framework that combines cell morphology attribute description and interpretable attribute matching is proposed for cervical cell classification.
🎯 What it does: Introducing the Reflect framework, using rectified flows for single-step correction of abnormal brain MRI in latent space and achieving unsupervised anomaly detection and localization.
Region-Based Text-Consistent Augmentation for Multimodal Medical Segmentation
Cai, Kunyan, Tan, Tao (Macao Polytechnic University)
CodeSegmentationConvolutional Neural NetworkLarge Language ModelMultimodalityBiomedical DataComputed Tomography
🎯 What it does: A region-based and text-consistent enhancement framework RBTCA is proposed, which maintains semantic consistency between multimodal medical images and text reports during data augmentation.
Grzeszczyk, Michal K. (Sano Centre for Computational Medicine), Sitek, Arkadiusz (Massachusetts General Hospital)
CodeExplainability and InterpretabilitySupervised Fine-TuningContrastive LearningMultimodalityTabularBiomedical DataMagnetic Resonance Imaging
🎯 What it does: This paper proposes a sparse interpretable regression scoring system called RegScore, and constructs a personalized linear regression (PLR) and personalized scoring (PRS) based on the TIP pre-trained model to integrate tabular and imaging data for predicting pulmonary artery pressure.
🎯 What it does: An Adaptive Low-Rank Parameter Tuning (ARENA) method is proposed, which improves the performance of traditional LoRA in medical image few-shot segmentation, capable of automatically adjusting low-rank parameters and enhancing segmentation quality without a validation set.
🎯 What it does: Enhancing the accuracy of medical image segmentation through a two-stage framework (image reconstruction and three-layer feature alignment),
CodeRestorationTransformerLarge Language ModelPrompt EngineeringImageBiomedical DataMagnetic Resonance Imaging
🎯 What it does: A cascade framework called ResMAP based on prompt retrieval is proposed for recovering magnetic resonance imaging (MRI) images containing various mixed artifacts;
ReSurgSAM2: Referring Segment Anything in Surgical Video via Credible Long-term Tracking
Liu, Haofeng (National University of Singapore), Jin, Yueming (University of Oxford)
CodeObject TrackingSegmentationTransformerVision Language ModelVideoMultimodality
🎯 What it does: A two-stage Surgical Referring Segmentation framework, ReSurgSAM2, is proposed, which first detects the target using text prompts and then performs long-term reliable tracking.
RetFiner: A Vision-Language Refinement Scheme for Retinal Foundation Models
Fecso, Ronald (Medical University of Vienna), Bogunović, Hrvoje (Medical University of Vienna)
CodeClassificationSegmentationTransformerVision Language ModelContrastive LearningImageTextBiomedical DataElectronic Health Records
🎯 What it does: A RetFiner visual-language refinement scheme is proposed, which refines existing retinal foundation models through self-supervised learning to enhance their semantic understanding and performance on downstream tasks.
CodeSegmentationGenerationData SynthesisDiffusion modelImageBiomedical Data
🎯 What it does: Utilizing a three-stage diffusion model (DDPM) to synthesize retinal OCT images with segmentation labels, thereby alleviating the problem of scarce labeled data and enhancing segmentation performance.
RetSTA: An LLM-Based Approach for Standardizing Clinical Fundus Image Reports
Cai, Jiushen, Li, Huiqi (Beijing Institute of Technology)
CodeTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: A bilingual retinal standard terminology was constructed, and two LLM models (RetSTA-7B-Zero and RetSTA-7B) were trained to achieve automatic standardization of fundus examination reports.
🎯 What it does: This paper proposes the Reverse Imaging method, which estimates the spin characteristics of cardiac MRI images through physics-guided reverse diffusion, enabling interpretable transformations between different sequences and zero-shot generalization.
🎯 What it does: This paper evaluates and proposes a highly generalizable bookmark supervision method by constructing a multi-task ScribbleBench benchmark.
🎯 What it does: This paper proposes a fully automated tile-level data refinement and sampling framework for visual foundation models (FM) in digital pathology: first, 350M tile embeddings are extracted using a pre-trained FM, and then balanced sampling is achieved using a hierarchical clustering tree; subsequently, a clustering-based hierarchical batch sampling is employed in self-supervised training to mitigate the negative impact of sample imbalance.
🎯 What it does: A multi-modal retinal image fusion framework RIFNet has been developed, which first synthesizes pseudo FA images using a dual-stream generative network, and then fuses CFP and FA using Viridis color encoding, ultimately generating high-quality, detail-rich fused images.
🎯 What it does: This paper proposes a fully automated myocardial scar segmentation pipeline based on YOLO and SAM, utilizing KL loss and extensive data augmentation to overcome issues of semi-automated annotation noise, data heterogeneity, and class imbalance.
🎯 What it does: A cross-population data augmentation framework has been developed, which incorporates fetus-specific image enhancements (such as fetal position compensation and amniotic fluid reconstruction) during training to achieve generalization of fetal pose estimation across different gestational weeks.
Robust Incomplete-Modality Alignment for Ophthalmic Disease Grading and Diagnosis via Labeled Optimal Transport
Yu, Qinkai (University of Exeter), Meng, Yanda (University Of Exeter)
CodeClassificationRecognitionOptimizationImageMultimodalityBiomedical Data
🎯 What it does: This paper proposes a multimodal alignment and fusion framework based on Labeled Optimal Transport (Labeled OT) to accurately grade and diagnose ophthalmic diseases even in the absence of retinal photos (fundus) or Optical Coherence Tomography (OCT).
Robust sensitivity control in digital pathology via tile score distribution matching
Pignet, Arthur (Owkin), Olivier, Antoine (Owkin)
CodeClassificationConvolutional Neural NetworkBiomedical Data
🎯 What it does: A threshold calibration method based on tile score distribution matching (Tile-Score Matching, TSM) is proposed to control the sensitivity of digital pathology WSI binary classification models in multi-instance learning (MIL) models.
Robust Sleep Stage Prediction from Electroencephalogram with Label Noise Using Multimodal Large Language Models
Qiu, Xihe (Shanghai University of Engineering Science), Tan, Xiaoyu (Shanghai University of Engineering Science)
CodeClassificationTransformerLarge Language ModelPrompt EngineeringMultimodalityTime SeriesBiomedical Data
🎯 What it does: A framework for sleep stage prediction based on a multimodal large language model (MLLM) is proposed, which addresses EEG label noise by using multi-perspective consistency to filter high-quality samples and employs a self-training loop to enhance model robustness.
RRG-DPO: Direct Preference Optimization for Clinically Accurate Radiology Report Generation
Liu, Hong (Eindhoven University of Technology), Wang, Liansheng (Shanghai Changhai Hospital)
CodeGenerationAnomaly DetectionOptimizationLarge Language ModelSupervised Fine-TuningTextBiomedical DataComputed TomographyElectronic Health RecordsRetrieval-Augmented Generation
🎯 What it does: A framework based on Direct Preference Optimization (RRG-DPO) is proposed, which fine-tunes the radiology report generation model using sub-preference loss, significantly reducing hallucination errors and improving clinical accuracy.
🎯 What it does: By constructing a self-supervised anomaly detection framework RSAD, brain disease diagnosis is achieved on fMRI data. The core idea is to first use a masked Transformer autoencoder to learn the reconstruction representation of healthy samples, and then utilize region-specific difference scores to determine anomalies in patients.
CodeImage TranslationGenerationDomain AdaptationRecurrent Neural NetworkGenerative Adversarial NetworkVideoBiomedical Data
🎯 What it does: A lightweight recursive temporal generative adversarial network (RT-GAN) is proposed to add temporal consistency to frame-based unsupervised domain translation models, addressing the inter-frame jitter problem in colonoscopy video analysis.
🎯 What it does: A structure-aware bidirectional proxy interaction network (SABPI-Net) was constructed for early ROP diagnosis based on color fundus images, achieving high-accuracy classification on a self-built dataset of over 170K images.
🎯 What it does: A unified framework for missing slice interpolation, SAGCNet, is proposed, which can restore complete cardiac magnetic resonance imaging (CMR) volumes under any missing slice scenario.
🎯 What it does: A hybrid network SAMASK-CLTR based on ResNet-50 and Transformer is proposed, achieving binary classification of benign and malignant tumors in ABUS images through mask prompts with 3D spatial position encoding.
SAMSA: Segment Anything Model enhanced with Spectral Angles for Hyperspectral Interactive Medical Image Segmentation
Roddan, Alfie (Imperial College London), Giannarou, Stamatia (Imperial College London)
CodeSegmentationImageBiomedical Data
🎯 What it does: Proposes the SAMSA framework, which combines the Segment Anything model with spectral angle similarity to achieve interactive segmentation of hyperspectral medical images.
🎯 What it does: This paper presents ScalpVision, a complete diagnostic system for scalp diseases that combines unlabeled hair segmentation and generative data augmentation.
Segmenting Vessels Encapsulating Tumor Clusters via Fine-Grained Visual Prompt
Yu, Jiahui (Zhejiang University), Xu, Yingke (University of British Columbia)
CodeSegmentationTransformerContrastive LearningImageBiomedical Data
🎯 What it does: This paper proposes a VPP2P framework based on pixel-level visual prompts, utilizing nuclear location priors to achieve semi-supervised fine-grained segmentation of VETC.
🎯 What it does: A structured causal model using a pre-trained segmenter for adversarial fine-tuning (Seg-CFT) is proposed, achieving local and controllable interventions on structure-specific attributes in medical images.
🎯 What it does: This paper proposes an end-to-end learning framework called LDOS for multi-organ segmentation on ultra-low dose PET (LDPET) without the need for CT assistance.
CodeClassificationRecognitionOptimizationConvolutional Neural NetworkMixture of ExpertsImageBiomedical Data
🎯 What it does: Using a self-propagating multi-task learning model to simultaneously predict 15 cardiovascular metabolic risk factors from fundus images.
🎯 What it does: A self-supervised volumetric super-resolution framework D2R is proposed, which first extracts biological structure priors using a 2D diffusion model, then generates pseudo high-resolution volumes, and finally learns stable structural transformations using a 3D VSR network to achieve axial resolution enhancement.
🎯 What it does: A large number of candidate events are generated using a legacy rule-based HFO detector, followed by learning event morphological representations through a self-supervised variational autoencoder (VAE), and weak labels are extracted using latent space clustering. Finally, a classifier is trained to accurately screen candidate events for pathological HFOs.
🎯 What it does: A Semantic Interpolation Diffusion Model (SIDM) is proposed, which generates diverse pairs of endoscopic images and masks by interpolating between lesion masks and background semantic labels to enhance the generalization ability of segmentation models.
Semantically Consistent Discrete Diffusion for 3D Biological Graph Modeling
Prabhakar, Chinmay (University of Zurich), Menze, Bjoern (University of Zurich)
CodeGenerationData SynthesisGraph Neural NetworkTransformerDiffusion modelGraphBiomedical Data
🎯 What it does: A semantic-consistent discrete diffusion model is designed to generate 3D biological graph networks that satisfy anatomical structure and edge label constraints.
🎯 What it does: A two-stage semi-supervised unpaired image translation framework is proposed, which generates high-quality, HU value-accurate CT images from CBCT while maintaining complete anatomical structure and high resolution.
🎯 What it does: This paper proposes a semi-supervised multimodal medical image segmentation framework that enhances segmentation accuracy using limited labeled data through multi-stage feature fusion, modality-aware enhancement, and contrastive mutual learning.
🎯 What it does: A semi-supervised surgical phase recognition framework based on video Transformer is proposed, integrating temporal consistency regularization and category prototype-based contrastive learning.
Separable tissue representations for attributable risk prediction
Wåhlstrand, Victor (Chalmers University of Technology), Häggström, Ida (Chalmers University of Technology)
CodeSegmentationExplainability and InterpretabilityRepresentation LearningTransformerAuto EncoderImageBiomedical DataComputed Tomography
🎯 What it does: Developed the STRAP framework, which utilizes variable-sized ROI masks to learn the separable representations of tissues such as bone, muscle, and fat in high-resolution CT images for fracture risk prediction and interpretability.
Sequence-Independent Continual Test-Time Adaptation with Mixture of Incremental Experts for Cross-Domain Segmentation
Xu, Dunyuan (Chinese University of Hong Kong), Heng, Pheng-Ann (Chinese University of Hong Kong)
CodeSegmentationDomain AdaptationConvolutional Neural NetworkMixture of ExpertsImageBiomedical DataMagnetic Resonance Imaging
🎯 What it does: An adaptive framework for continuous testing based on Mixture of Incremental Experts (MoIE) is proposed for cross-domain medical image segmentation, which can maintain stable performance in the case of randomly arriving samples.
ShareLink: Neuro-Inspired EEG-based Cross-Subject Emotion Recognition via Shared Bi-hemisphere
Huang, Zixuan (Chinese Academy of Sciences), Miao, Fen (University of Electronic Science and Technology of China)
CodeRecognitionGraph Neural NetworkMixture of ExpertsTime SeriesBiomedical Data
🎯 What it does: The ShareLink model is proposed to achieve cross-subject EEG emotion recognition by enhancing recognition performance through the alignment and interaction of features from the left and right hemispheres.
Shuffle-Diversity Collaborative Federated Learning for Imbalanced Medical Image Analysis
Gao, Wenpeng (South China Agricultural University), Fan, Xiaomao (Shenzhen Technology University)
CodeClassificationFederated LearningConvolutional Neural NetworkImageBiomedical Data
🎯 What it does: Proposes the FedSDC joint network, which addresses sample imbalance and heterogeneity in medical image classification through shared Body and multi-head Head in federated learning, using Shuffle-Diversity collaboration.
Sim-to-Real Transformer-Based Shape Reconstruction for Automated Orthopedic Fracture Reduction Planning
Yibulayimu, Sutuke (Beihang University), Wang, Yu (Capital Medical University)
CodeTransformerBiomedical DataComputed Tomography
🎯 What it does: Proposes an automated fracture reduction planning framework based on Transformer for fracture shape recovery and recursive registration.
🎯 What it does: Through the SimCroP framework, which combines cross-granularity pre-training with similarity-driven alignment, joint pre-training of chest CT and corresponding reports is conducted, significantly improving classification and segmentation performance under multi-task settings.
🎯 What it does: For the test-time domain adaptation of single medical images, a multi-view collaborative training method guided by uncertainty is proposed, allowing the model to perform adaptive learning on a single target image only during inference, thus achieving 3D breast MRI tumor segmentation.
🎯 What it does: A multi-modal multi-scale fusion network (MMCAF-Net) aimed at classifying lung diseases is proposed, which simultaneously processes 3D medical images and clinical tabular data, focusing on capturing small lesion information.
🎯 What it does: A multi-teacher self-distillation framework is proposed for surgical video analysis, which selects multiple rounds of teacher models with different random seeds within the 'practical zone' during the training process and averages their soft labels to enhance the performance of the student model; subsequently, a temporal decoder is added to extract spatiotemporal information.
🎯 What it does: A source-free domain adaptation framework is proposed, utilizing a teacher-student model and contrastive learning to improve cross-modal cardiac image segmentation.
🎯 What it does: This paper proposes a Sparse Optical Doppler Tomography Reconstruction Network based on an Alternative State Space Model and Attention Mechanism (ASSAN), which can quickly generate high-quality B-Scan images with fewer A-Scan data.
🎯 What it does: A fully automated framework for horizontal pose correction (Sparse-XM) is proposed, utilizing intraoperative stereo vision (iSV) images for spinal bone surface segmentation, and achieving spinal pose adjustment through alignment of the bone surface with preoperative CT (pCT).
🎯 What it does: The SA-Net framework is proposed, utilizing the Segment Anything Model (SAM) with cross-supervision from 2D and 3D networks, combined with Bidirectional Graph Convolution (BGC), Multi-Scale Attention (MSA), and a Recovery Module (RM) to achieve sparse annotation in medical image segmentation.
🎯 What it does: A meta-learning based semi-supervised domain adaptation framework is proposed to learn noise features and perform denoising from sparsely labeled fMRI data.
🎯 What it does: The STAR framework is proposed, combining semi-supervised learning and active learning, generating pseudo-labels and queries between 2D slices of 3D medical images through spatial cross-attention, significantly reducing annotation costs.
🎯 What it does: An efficient segmentation framework for neonatal hypoxic-ischemic encephalopathy (HIE) lesions based on a 2D UNet++ network is proposed, utilizing three-channel input (ADC, ZADC, and thresholded ZADC) combined with scSE attention and TLHF loss for precise segmentation.
Spatial regularisation for improved accuracy and interpretability in keypoint-based registration
Billot, Benjamin (Inria), Golland, Polina (Boston Children's Hospital)
CodeSegmentationExplainability and InterpretabilityConvolutional Neural NetworkTime SeriesBiomedical DataMagnetic Resonance ImagingAlzheimer's Disease
🎯 What it does: A triple spatial regularization loss is proposed, significantly improving the performance of unsupervised keypoint detection-based registration methods in terms of accuracy and interpretability.
🎯 What it does: A spatiotemporal memory filtering network based on the Segment Anything model (STMFSAM) is designed and implemented for lesion segmentation in breast ultrasound videos.
Spatially Gene Expression Prediction using Dual-Scale Contrastive Learning
Qu, Mingcheng (Harbin Institute of Technology), Fan, Lei (University of New South Wales)
CodeSegmentationRepresentation LearningGraph Neural NetworkContrastive LearningImageMultimodalityBiomedical Data
🎯 What it does: A dual-scale contrastive learning framework NH²2ST is proposed, which predicts spatial gene expression from whole slide images by combining cross-attention, contrastive learning, and hypergraph learning in the query branch and neighborhood branch.